Edge AI · Physical AI · On-device GenAI · 2026 field guide
Top 100 Embedded AI Projects for Embedded Engineers
A current, buildable catalogue of on-device intelligence for 2026, from coin-cell wake words to on-device language models and physical-AI robots. Every project explained, with today’s hardware and an estimated prototype cost.
Why run AI on the device itself?
For decades, “adding intelligence” to a product meant streaming data to a server and waiting for an answer. Embedded AI, also called edge AI or TinyML, inverts that model: the neural network runs on the microcontroller or single-board computer inside the product. Nothing has to leave the device. The four reasons this matters are consistent across every project below.
Latency. A balancing robot or a collision warning cannot afford a network round trip. On-device inference completes in milliseconds, inside the control loop, every time.
Privacy. Cameras, microphones and health sensors capture deeply personal data. When the model runs locally and only a decision leaves the device, the raw signal never becomes a liability.
Reliability. Factories, farms and vehicles often have poor or no connectivity. Edge AI keeps working when the network is down, which for security devices is a feature, not a footnote.
Cost & energy. No cloud compute bill, and a well-designed pipeline sleeps between events, letting a battery node run for months on a coin cell.
The ten families of embedded AI
The 100 projects are organized into ten categories, each with its own signal-flow pattern and its own dominant constraint, energy for field nodes, latency for robots and vehicles, privacy for wearables and cameras. This overview shows the size and cost spread of each family; the detailed sections follow.
| # | Category | Projects | Cost range |
|---|---|---|---|
| 01 | Computer Vision on the Edge | 10 | $30–$330 |
| 02 | Audio, Speech & On-Device Voice | 10 | $22–$160 |
| 03 | Predictive Maintenance & Industrial IoT | 10 | $40–$75 |
| 04 | Robotics & Physical AI | 10 | $60–$500 |
| 05 | Health, Fitness & Wearables | 10 | $30–$70 |
| 06 | Smart Agriculture & Environment | 10 | $45–$165 |
| 07 | Smart Home & On-Device Assistants | 10 | $35–$200 |
| 08 | Automotive & Mobility | 10 | $45–$330 |
| 09 | Security, Safety & Surveillance | 10 | $35–$320 |
| 10 | On-Device Generative AI & Edge MLOps | 10 | $30–$170 |
Cost and difficulty at a glance
Average hardware cost climbs with sensing richness: audio and inertial nodes are cheapest, camera-based vision and robotics are dearest because they need more capable compute. The difficulty mix is deliberately balanced so there is an on-ramp at every level.
Computer Vision on the Edge
In 2026 edge vision has moved past single-class detectors: cheap NPU boards like the Raspberry Pi 5 AI Kit and Jetson Orin Nano Super now run open-vocabulary detection and small vision-language models (VLMs) that describe a scene in words, all on-device. These projects still teach the core skills, quantization, camera interfacing, region-of-interest logic, but the ceiling is far higher than it was two years ago.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | Person Detection Doorbell | ESP32-P4 + OV5640 camera | Beginner | $30 |
| 02 | VLM Scene-Describer for the Blind NEW | Raspberry Pi 5 + AI Kit + camera | Advanced | $170 |
| 03 | Real-Time Face Mask & PPE Checker | Raspberry Pi 5 + AI Kit + camera | Intermediate | $160 |
| 04 | Analog Meter Reader | ESP32-P4 + camera + LoRa | Intermediate | $45 |
| 05 | Object Counting Conveyor Camera | Jetson Orin Nano Super + industrial camera | Advanced | $330 |
| 06 | Gesture-Controlled Interface | Raspberry Pi 5 + camera | Intermediate | $90 |
| 07 | License Plate Recognition Gate | Jetson Orin Nano Super + IP camera | Advanced | $320 |
| 08 | Solar Wildlife Camera Trap | ESP32-P4 camera + solar + LoRa | Intermediate | $65 |
| 09 | Anomaly-Based 3D Print Monitor | Raspberry Pi 5 + camera | Intermediate | $100 |
| 10 | Open-Vocabulary Shelf Monitor NEW | Raspberry Pi 5 + AI Kit + camera | Advanced | $165 |
Person Detection Doorbell
A battery doorbell that wakes on motion and runs a tiny INT8 detector locally before ever streaming video, so it notifies you only when a real person is present instead of firing on every shadow. Modern ESP32-P4 silicon adds an AI-friendly instruction set that lets the whole capture-and-classify pipeline finish in tens of milliseconds inside a few hundred kilobytes of memory, while the camera and inference core sleep between events to stretch battery life to months. It remains the canonical first serious edge-vision build and teaches quantization, frame capture, and aggressive duty cycling.
VLM Scene-Describer for the Blind
A wearable or handheld camera that captures a frame and speaks a natural-language description of the scene, read the sign, count the steps, what is on the table, using a small on-device vision-language model so nothing is sent to a server. This is one of the defining 2026 edge builds: a compact VLM turns pixels into a sentence locally, giving genuinely useful assistive output with full privacy. Engineers learn to run a quantized multimodal transformer on an NPU board and to fuse vision with local text-to-speech.
Real-Time Face Mask & PPE Checker
An entrance unit that detects each face or worker and checks for required items, mask, hard hat, goggles, driving a simple pass/fail indicator and logging aggregate counts without storing identifiable images. A detector finds people, then attribute classifiers judge each item, all fused into one graph running on a Hailo NPU at a steady frame rate. Because faces are processed and discarded on-device, it sidesteps most privacy concerns. Engineers learn cascaded inference, region-of-interest cropping, and keeping several small models cooperating in one latency budget.
Analog Meter Reader
A clamp-on module that photographs an analog utility meter's digit wheels and reads the value with a compact CNN trained on odometer and seven-segment fonts, then reports over long-range radio a few times a day so old mechanical meters join a modern telemetry network. The hard part is robust preprocessing, correcting glare, tilt, and half-rolled digits under changing daylight, not the classifier itself. It is an excellent lesson in real-world image normalization, where the model is easy and the lighting is the enemy.
Object Counting Conveyor Camera
An overhead camera on a production line that counts and classifies discrete parts as they pass, feeding tallies to a factory dashboard and continuing to count even if the plant network drops. A segmentation-plus-tracking pipeline on an Orin Nano Super handles overlapping and touching items, and a line-crossing tracker prevents double counting when parts jitter. Engineers confront motion blur, exposure tuning for fast belts, and the surprisingly tricky logic of counting objects that touch, split, and occlude in a continuous stream.
Gesture-Controlled Interface
A desk or wearable camera that recognizes a small vocabulary of hand gestures, swipe, point, pinch, to control slides, media, or a smart display with no physical remote. A tiny keypoint model extracts hand landmarks, then a lightweight temporal classifier maps landmark sequences to gestures so a swipe is distinguished from a wave. Keeping false triggers low while the hand rests is the real problem, solved with confidence gating and a short debounce window. It teaches keypoint estimation and temporal smoothing on constrained hardware.
License Plate Recognition Gate
A parking gate that captures approaching vehicles, locates the plate, rectifies it, and runs on-device OCR to open the barrier only for vehicles on a locally stored allowlist, keeping the resident database private and the gate responsive without a network round trip. Detection and recognition run in two stages on an NPU board, with perspective correction to handle angle and distance. Engineers learn multi-stage detection, image rectification, and building a fast on-device lookup against a stored list.
Solar Wildlife Camera Trap
A solar trail camera that classifies captured animals into species on the spot, so researchers receive labeled sightings instead of thousands of empty frames, flagging rare or low-confidence images for human review. A quantized classifier fires only after a motion wake, and the extreme constraint is energy, the device must survive weeks in the field on a small panel and battery, so inference is rationed carefully. It is a masterclass in energy-aware scheduling and handling long-tailed, imbalanced class distributions.
Anomaly-Based 3D Print Monitor
A camera watching a 3D printer that detects failures, spaghetti, layer shifts, warping, and pauses the print before wasting hours of filament, integrating with the printer firmware over a network hook so it can act, not just alert. An anomaly-style model trained mostly on healthy prints flags deviations rather than memorizing every failure mode, which generalizes to new models. The project teaches one-class anomaly detection and closing the loop from perception to physical control.
Open-Vocabulary Shelf Monitor
A fixed camera that estimates how full a shelf is and flags gaps for restocking, and, using an open-vocabulary detector, can be told to find new products by name without retraining, all on-device so no customer footage leaves the store. It outputs a fill ratio per shelf region rather than tracking individual SKUs, staying robust to packaging changes. Engineers learn segmentation plus modern open-vocabulary detection on edge hardware and turning pixel masks into an actionable out-of-stock metric.
Audio, Speech & On-Device Voice
Audio is still the most power-efficient always-on modality, but 2026 raised the bar: compact speech-to-text models now run entirely on-device, so a microcontroller can not only spot a wake word but transcribe a full command without the cloud. These projects span the whole stack, from coin-cell keyword spotting to local Whisper-class transcription and voice interfaces that keep every word private.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | Custom Wake-Word Engine | Arduino Nano ESP32 | Beginner | $35 |
| 02 | Offline Speech-to-Text Note Taker NEW | Raspberry Pi 5 + AI Kit + mic | Advanced | $160 |
| 03 | Machine Fault Listening Node | STM32N6 (Neural-ART NPU) + MEMS mic | Intermediate | $60 |
| 04 | Fully-Offline Voice Command Hub NEW | ESP32-P4 + microphone | Intermediate | $40 |
| 05 | Gunshot & Glass-Break Alarm | Nordic nRF54L15 + MEMS mic | Intermediate | $40 |
| 06 | Baby Cry & Nursery Monitor | ESP32-S3 + I2S MEMS mic | Beginner | $22 |
| 07 | Cough & Respiratory Screener | Nordic nRF54L15 + MEMS mic | Intermediate | $45 |
| 08 | Bird & Wildlife Sound Logger | Raspberry Pi 5 + USB mic + solar | Intermediate | $130 |
| 09 | Snore & Sleep-Sound Analyzer | ESP32-S3 + MEMS mic | Beginner | $30 |
| 10 | Industrial Alarm-Tone Recognizer | STM32N6 (Neural-ART NPU) + MEMS mic | Intermediate | $55 |
Custom Wake-Word Engine
An always-listening module that detects a chosen wake phrase and only then powers up a heavier downstream system, the pattern behind every voice assistant. Audio is framed, converted to a compact spectral feature, and streamed through a small classifier that must run continuously at microamp-level average current. The defining challenge is the trade-off between false accepts and false rejects, too sensitive and it wakes at noise, too strict and it ignores the user. Engineers learn feature extraction, streaming inference, and rigorous threshold tuning against realistic negative audio.
Offline Speech-to-Text Note Taker
A pocket device that transcribes spoken notes to text entirely on-device using a compact speech-recognition model, so sensitive dictation never touches a cloud service. Running a quantized transformer-based recognizer on an NPU-class board turns hours of audio into searchable text locally, a capability that only became practical on cheap hardware recently. Engineers learn to deploy a small automatic-speech-recognition model, manage streaming audio buffers, and handle the memory footprint of a sequence model on the edge.
Machine Fault Listening Node
A sensor that clamps to a motor or pump and continuously listens for the acoustic signature of impending failure, bearing whine, cavitation, imbalance, before it is visible in vibration, sending alerts over an industrial bus while raw audio stays on the node. The model learns the healthy sound profile and scores deviation, so it needs little or no failure data to be useful. It teaches acoustic anomaly detection and the maintenance reality that normal is abundant and faults are rare.
Fully-Offline Voice Command Hub
A voice interface that recognizes a flexible set of commands and controls appliances with no cloud dependency and no privacy exposure, going beyond a fixed keyword list to a small on-device recognizer that understands natural phrasing. Local processing is the selling point for devices in bedrooms and bathrooms. Engineers learn small-vocabulary speech recognition, intent parsing, and mapping recognized phrases to safe, debounced actuator actions.
Gunshot & Glass-Break Alarm
A safety node that recognizes sharp impulsive events, glass breaking, a slammed door, an alarm tone, and triggers a response within a fraction of a second, firing even if connectivity is severed. Impulsive sounds are hard because they are brief and easily confused, so the model leans on both spectral shape and temporal envelope. The project explores transient event detection and the careful false-alarm engineering that safety devices demand.
Baby Cry & Nursery Monitor
A monitor that distinguishes an infant's cry from ambient household sounds and sends a discreet alert without streaming any audio off the device, working even with the home network down. A short spectrogram window feeds a compact classifier trained to separate crying from talking, music, and appliance noise. It is a gentle introduction to environmental sound classification and to curating a negative dataset that reflects a real home.
Cough & Respiratory Screener
A wearable or bedside monitor that counts coughs and characterizes them over time, giving clinicians a passive symptom trend without manual logging and keeping sensitive audio on-device. Cough events are detected by a compact classifier tuned to reject speech and throat clearing, and counts are aggregated into a daily record. Engineers learn medical-adjacent sound classification and the ethical weight of on-device processing for sensitive signals.
Bird & Wildlife Sound Logger
A field logger that recognizes bird calls by species and builds a local biodiversity record, running for weeks on batteries in remote habitats. Calls are converted to spectrograms and classified against a regional species set, with unknowns logged for later expert review. The energy and storage constraints mirror the camera trap but in the audio domain. It teaches fine-grained audio classification and dataset curation for many acoustically similar classes.
Snore & Sleep-Sound Analyzer
A bedside device that classifies nighttime sounds, snoring, teeth grinding, restlessness, into a morning sleep-quality summary, all processed locally overnight without storing raw recordings. A lightweight classifier runs on rolling audio windows and accumulates event statistics. The overnight, low-power, private nature of the task makes edge processing the natural fit. Engineers learn long-duration continuous inference and summarizing hours of events into a digestible report.
Industrial Alarm-Tone Recognizer
A node in a noisy plant that recognizes standardized machine alarm tones and forklift horns amid heavy background noise, alerting workers or logging events without the latency and dropout risk of streaming audio to a server. Because alarm tones are structured, the model keys on their spectral pattern while rejecting the surrounding din. The project focuses on noise-robust classification and exploiting known structure in the target signal.
Predictive Maintenance & Industrial IoT
Vibration, current, temperature and pressure signatures reveal machine health long before failure, and this remains where TinyML delivers the clearest financial return, a few dollars of MEMS sensor against avoided downtime. The 2026 change is silicon: MCUs like the STM32N6 now embed a dedicated Neural-ART NPU, so richer models run on the sensor node itself without a separate accelerator.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | Bearing Failure Predictor | STM32N6 (Neural-ART NPU) + MEMS accelerometer | Intermediate | $55 |
| 02 | Motor Current Signature Analyzer | ESP32-S3 + current transformer | Intermediate | $40 |
| 03 | Pump Cavitation Detector | STM32N6 + pressure + vibration sensors | Advanced | $70 |
| 04 | Compressed-Air Leak Locator | STM32N6 + ultrasonic mic | Advanced | $70 |
| 05 | Transformer Health Monitor | ESP32-S3 + temp + mic sensors | Intermediate | $60 |
| 06 | Conveyor Belt Slip & Wear Sensor | STM32N6 + encoder + accelerometer | Intermediate | $55 |
| 07 | HVAC Efficiency Optimizer | ESP32-S3 + temp/humidity sensors | Intermediate | $45 |
| 08 | Tank Level & Overflow Predictor | ESP32-S3 + ultrasonic level sensor | Beginner | $40 |
| 09 | Industrial Robot Joint Monitor | STM32N6 + current + IMU sensors | Advanced | $75 |
| 10 | Steam-Trap Failure Detector | STM32N6 + temp + acoustic sensors | Intermediate | $60 |
Bearing Failure Predictor
A vibration node bolted to a motor housing that learns the healthy vibration spectrum and scores how far current behavior has drifted, warning of bearing wear weeks ahead of failure, and it works on machines with no reliable network because it decides locally. A three-axis accelerometer feeds spectral features into an anomaly model, so no labeled failure data is required to start. This is the flagship predictive-maintenance build and teaches vibration feature engineering and one-class modeling.
Motor Current Signature Analyzer
A clamp on a motor's supply line that infers mechanical faults, misalignment, broken rotor bars, load imbalance, purely from the current waveform, a technique needing no sensor inside the machine, which makes retrofitting old equipment trivial. Harmonic features are extracted and classified against known fault signatures. Engineers learn that mechanical problems leave electrical fingerprints, and how to extract them from a current-transformer signal.
Pump Cavitation Detector
A node combining pressure and vibration to detect cavitation, the damaging formation and collapse of bubbles that erodes impellers, before it destroys the pump, keeping critical pumps protected regardless of network state. A small model fuses two sensor streams and flags the onset signature so operators can adjust flow. The project teaches sensor fusion and detecting a physical process by its combined mechanical and hydraulic symptoms.
Compressed-Air Leak Locator
A handheld or fixed ultrasonic detector that hears the high-frequency hiss of compressed-air leaks, an enormous hidden energy cost in factories, and classifies leak severity, giving an immediate leak/no-leak readout for a walking survey. An ultrasonic microphone shifts the inaudible leak into a band a small classifier can grade. It introduces ultrasonic sensing and the framing of AI as an energy-efficiency tool.
Transformer Health Monitor
A monitor on an electrical transformer tracking temperature, load, and acoustic hum to predict insulation aging and overload risk, with on-device decision-making that avoids reliance on backhaul from remote substations. A model correlates the multivariate trend with degradation and raises graded alerts. Engineers learn multivariate time-series modeling and mapping slow-moving trends to remaining useful life.
Conveyor Belt Slip & Wear Sensor
A node measuring belt speed and vibration to detect slippage, misalignment, and wear on material-handling conveyors, with local processing that suits the long, distributed layout of conveyor systems. Encoder and accelerometer data feed a model that separates normal load variation from developing faults. The project covers combining rotational and vibration sensing to monitor a large mechanical system from one point.
HVAC Efficiency Optimizer
An embedded controller that learns a building's thermal response and predicts the minimum runtime needed to hit setpoints, trimming energy without sacrificing comfort, running on the equipment so control stays fast and resilient. A small predictive model anticipates heating and cooling demand from temperature, occupancy, and time. Engineers learn predictive control and modeling a slow physical system with a compact network.
Tank Level & Overflow Predictor
A monitor that forecasts when a liquid tank will fill or empty from level trends and flow, preventing overflows and dry-running pumps, with on-device prediction that matters for remote tanks with intermittent connectivity. A lightweight forecaster projects the level curve forward and triggers action with lead time. It teaches short-horizon time-series forecasting and acting on a prediction rather than a threshold.
Industrial Robot Joint Monitor
A node on a robot arm that watches joint current and vibration to detect gearbox wear and backlash before positioning accuracy degrades, adding a health layer without touching the robot's own control loop. Per-joint models learn each axis's healthy signature during normal motion cycles. Engineers learn per-axis modeling and non-intrusive monitoring of a complex articulated machine.
Steam-Trap Failure Detector
A clamp-on node that determines whether a steam trap is working, stuck open (wasting steam), or stuck closed (risking water hammer), from its temperature and acoustic pattern, scaling far better than manual inspection across a large distributed steam system. A compact classifier grades trap state from the fused signal. The project pairs thermal and acoustic sensing to diagnose a device by its behavior.
Robotics & Physical AI
Physical AI, robots that perceive, reason and act in the real world, is the headline theme of 2026, driven by new compute like NVIDIA Jetson Thor and Qualcomm Dragonwing aimed squarely at autonomous mobile robots and humanoids. On-device intelligence is non-negotiable here: a cloud round trip is far too slow to balance, grasp or avoid, so these projects run perception and policy inside a millisecond control loop.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | Self-Balancing Robot with Learned Control | ESP32-S3 + MPU-6050 IMU + motors | Intermediate | $60 |
| 02 | Vision-Language-Action Manipulator NEW | Jetson Orin Nano Super + camera + servo arm | Advanced | $420 |
| 03 | Object-Sorting Robotic Arm | Raspberry Pi 5 + camera + servo arm | Advanced | $180 |
| 04 | Autonomous Mobile Robot (AMR) NEW | Qualcomm Dragonwing dev board + depth camera + chassis | Advanced | $500 |
| 05 | Gesture-Piloted Drone | Raspberry Pi 5 + camera + drone kit | Advanced | $220 |
| 06 | Adaptive Gripper Force Controller | STM32N6 + force sensors + servo gripper | Advanced | $120 |
| 07 | Terrain-Adaptive Quadruped NEW | Jetson Orin Nano Super + servos + IMU | Advanced | $340 |
| 08 | Visual Servoing Tracker | Raspberry Pi 5 + camera + pan-tilt servos | Intermediate | $110 |
| 09 | SLAM Indoor Mapper | Jetson Orin Nano Super + LiDAR + chassis | Advanced | $380 |
| 10 | Teleoperation Latency Assistant | STM32N6 + haptic sensors + motors | Advanced | $140 |
Self-Balancing Robot with Learned Control
A two-wheeled robot that stays upright using a policy trained to balance, going beyond hand-tuned PID by learning to reject disturbances it was never explicitly programmed for. An IMU feeds the state estimate, and the control loop must run at high frequency with deterministic timing, so on-device inference is non-negotiable because any latency topples the robot. It is a superb introduction to control-oriented ML and hard real-time inference budgets.
Vision-Language-Action Manipulator
A desktop arm you can instruct in plain language, pick up the red block and put it in the bowl, using a compact vision-language-action policy that maps a camera view and a text command directly to arm motions. This is the frontier 2026 robotics build: instead of scripting every pick, a single learned model generalizes to new objects and phrasings, running on a physical-AI compute board for low latency. Engineers learn multimodal policy deployment and grounding language in physical action.
Object-Sorting Robotic Arm
A desktop arm that identifies and sorts items by type or color into bins, closing the loop from camera to gripper with perception run locally so pick cycles stay fast. A vision model localizes and classifies the object, and inverse kinematics places the gripper. The project unites vision, coordinate transforms, and actuation into a complete pick-and-place system, the essence of industrial automation in miniature.
Autonomous Mobile Robot (AMR)
A rover that navigates cluttered spaces using depth and multi-sensor perception with a learned avoidance policy, handling obstacles a fixed rule set would miss, and running fully autonomously where no operator or network is available. Sensor data feeds a model that outputs safe headings in real time on physical-AI-class compute. Engineers learn reactive navigation and fusing range sensing with a learned decision layer, the building block of warehouse and service robots.
Gesture-Piloted Drone
A small drone that responds to a pilot's hand gestures captured by an onboard or ground camera, translating poses into flight commands with strict safety limits, where latency and reliability are safety-critical and mandate local inference. A keypoint model reads the gesture and a command mapper enforces envelope constraints. The project covers real-time pose interpretation and safe command arbitration for a flying platform.
Adaptive Gripper Force Controller
A robotic gripper that learns how much force to apply to hold objects of varying fragility without crushing or dropping them, from tactile and current feedback, adapting to unseen objects with an on-device control loop that stays responsive. A small model maps grip feedback to a target force. Engineers learn tactile sensing and force control framed as a learned regression problem.
Terrain-Adaptive Quadruped
A four-legged robot that learns stable walking gaits and adapts stride to terrain, going beyond fixed gait tables, with a policy that maps body state to leg commands at a high control rate inside a tight sensorimotor loop that demands local inference. This ambitious physical-AI build teaches legged locomotion, high-frequency control, and deploying a learned policy on constrained hardware.
Visual Servoing Tracker
A pan-tilt camera rig that keeps a moving target centered in frame using a learned visual controller, a building block for tracking cameras and turrets, with local processing that avoids the lag that would make tracking jittery. The model regresses correction commands from the target's image position, closing a smooth tracking loop. Engineers learn visual servoing and continuous control from image feedback.
SLAM Indoor Mapper
A small robot that builds a map of a room while localizing within it, using lightweight learned features to make simultaneous localization and mapping tractable on a low-power board and enabling autonomy without external infrastructure. Range and odometry data are fused into a consistent map. The project introduces the perennial SLAM problem in an embedded-friendly form.
Teleoperation Latency Assistant
A teleoperated manipulator whose embedded controller predicts operator intent and smooths commands, compensating for network jitter so remote control feels immediate, with edge intelligence bridging the human and the delayed link. A small predictor anticipates the next motion and fills gaps during lag. Engineers learn intent prediction and latency compensation for shared human-robot control.
Health, Fitness & Wearables
Wearables generate continuous, deeply personal biosignals that are best analyzed on-body, and 2026's ultra-low-power parts like the Nordic nRF54 family let richer models run for days on a coin cell. Edge AI here means keeping health data private by never transmitting the raw signal, and inferring meaningful states from noisy sensors strapped to a moving human.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | Fall Detection Wearable | Nordic nRF54L15 + IMU | Intermediate | $40 |
| 02 | Atrial Fibrillation Screener | Optical heart-rate sensor + nRF54L15 | Advanced | $55 |
| 03 | Activity & Exercise Recognizer | Arduino Nano ESP32 + IMU | Beginner | $40 |
| 04 | Sleep-Stage Estimator | Optical HR + IMU + nRF54L15 | Advanced | $50 |
| 05 | Posture & Ergonomics Coach | IMU + nRF54L15 + haptic motor | Beginner | $40 |
| 06 | Stress & HRV Monitor | HR + GSR sensors + nRF54L15 | Advanced | $55 |
| 07 | Sweat & Hydration Patch | Sweat sensor + nRF54L15 | Advanced | $65 |
| 08 | Tremor & Movement Logger | IMU + nRF54L15 | Intermediate | $45 |
| 09 | Smart Insole Gait Analyzer | Pressure array + nRF54L15 | Advanced | $70 |
| 10 | UV Exposure & Skin-Safety Band | UV sensor + nRF54L15 | Beginner | $30 |
Fall Detection Wearable
A wrist or waist device that distinguishes a genuine fall from everyday motion, sitting hard, jumping, dropping the device, and triggers an alert even without a phone nearby, which is crucial for elderly independence. An IMU stream feeds a compact classifier that must catch real falls while almost never crying wolf. It is a canonical wearable ML build and a lesson in tuning for rare, high-stakes events.
Atrial Fibrillation Screener
A wearable that analyzes heart-rhythm intervals to flag possible atrial fibrillation and prompt the user to seek a clinical ECG, processing cardiac data on-device to keep it private and work offline. A small model examines beat-to-beat variability for the irregular pattern. Engineers learn physiological time-series analysis and the careful, non-diagnostic framing a screening tool requires.
Activity & Exercise Recognizer
A fitness band that classifies activity type, walking, running, cycling, rowing, and counts repetitions, replacing manual workout logging, with everything running on the band for instant feedback and long battery life. Motion features feed a multi-class classifier and a rep-counting layer. The project teaches human-activity recognition, the classic accelerometer ML problem, and repetition detection from periodic motion.
Sleep-Stage Estimator
A wearable that estimates sleep stages, light, deep, REM, from motion and heart rate, producing a morning report without a clinical lab, where long, low-power, private overnight operation makes on-device processing the right choice. A compact model maps overnight signal patterns to stages. Engineers learn multimodal physiological modeling and long-duration inference on a battery budget.
Posture & Ergonomics Coach
A clip-on or garment sensor that detects slouching and prolonged static posture and nudges the wearer to correct it, with on-device feedback that is immediate and keeps posture data private. Orientation features feed a classifier that recognizes poor posture and stillness. The project covers orientation sensing and turning a classification into gentle, well-timed behavioral prompts.
Stress & HRV Monitor
A wearable that estimates stress from heart-rate variability and skin response, surfacing trends and suggesting breathing breaks, with sensitive biosignals kept on the device by design. A small model maps physiological features to a stress index. Engineers learn HRV feature extraction and translating a physiological estimate into an actionable wellness signal.
Sweat & Hydration Patch
A skin patch that estimates hydration and electrolyte loss during exercise from sweat-sensing electrodes, guiding fluid intake with on-body analysis that gives athletes real-time guidance. A compact model interprets the electrochemical signal into hydration status. The project introduces electrochemical sensing and modeling a chemical measurement on tiny hardware.
Tremor & Movement Logger
A wearable that quantifies tremor and movement patterns to help track conditions like Parkinson's, giving clinicians objective daily data through private, continuous, on-device logging. A model characterizes tremor frequency and severity from motion. Engineers learn frequency-domain motion analysis and building a longitudinal health record.
Smart Insole Gait Analyzer
A pressure-sensing insole that analyzes gait and balance, flagging asymmetry and fall risk for rehabilitation and sports, with on-device analysis that makes the insole a self-contained lab. A model reads the pressure map and stride timing to characterize gait. The project covers pressure-array sensing and biomechanical modeling in an extremely space-constrained form.
UV Exposure & Skin-Safety Band
A wearable that tracks UV exposure and models cumulative skin dose against skin type, warning before overexposure, with on-device tracking that gives real-time private guidance outdoors. A small model integrates UV readings into a personalized risk estimate. Engineers learn environmental sensing and personalized dose modeling from a simple sensor stream.
Smart Agriculture & Environment
Farms and remote environments are the natural home of low-power edge AI: connectivity is poor, power is scarce, and decisions are local. In 2026, solar nodes pair capable NPU boards with long-range radio to classify pests, grade crops and monitor ecosystems for months, sending only distilled insights instead of raw data, and increasingly using VLMs to describe field conditions in plain language.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | Crop Disease Leaf Scanner | Raspberry Pi 5 + camera | Intermediate | $110 |
| 02 | Solar Smart Pest Trap | ESP32-P4 camera + solar + LoRa | Intermediate | $65 |
| 03 | Livestock Behavior Monitor | IMU + nRF54L15 + LoRa | Intermediate | $50 |
| 04 | Soil Condition Advisor | ESP32-S3 + soil sensors + solar | Beginner | $45 |
| 05 | Fruit Ripeness & Grading Camera | Raspberry Pi 5 + camera | Intermediate | $110 |
| 06 | Precision Weed Sprayer NEW | Raspberry Pi 5 + AI Kit + camera | Advanced | $165 |
| 07 | Microclimate Frost Predictor | ESP32-S3 + weather sensors + solar | Intermediate | $60 |
| 08 | Water Quality Sentinel | ESP32-S3 + water sensors + solar | Intermediate | $70 |
| 09 | Greenhouse Climate Optimizer | ESP32-S3 + sensors + relays | Intermediate | $55 |
| 10 | Forest-Fire Early-Warning Node | ESP32-S3 + gas/temp sensors + solar + LoRa | Intermediate | $65 |
Crop Disease Leaf Scanner
A handheld or fixed camera that diagnoses plant disease from leaf images in the field, giving farmers an instant read where no agronomist or internet is available, which makes offline operation the whole point. A classifier trained on healthy and diseased leaves names the likely condition. Engineers learn fine-grained visual classification and building for genuinely disconnected environments.
Solar Smart Pest Trap
A solar trap that photographs captured insects, classifies species, and counts pest pressure, replacing weekly manual inspection and running for a full season on energy-frugal on-device inference. A compact model identifies target pests and logs counts over long-range radio. The project teaches insect classification and season-long low-power field operation.
Livestock Behavior Monitor
An ear-tag or collar sensor that classifies animal behavior, grazing, ruminating, resting, lameness, to flag health and estrus early, with on-animal low-power inference that scales across a whole herd. Motion features feed a behavior classifier on the tag itself. Engineers learn animal-activity recognition and inferring health from behavior patterns.
Soil Condition Advisor
A buried node that reads soil moisture, temperature, and nutrients and recommends irrigation and fertilization timing, learning the field's local response and making autonomous local decisions suited to fields with no network. A small model turns raw soil signals into actionable guidance. The project covers multi-sensor soil analysis and translating measurements into farm actions.
Fruit Ripeness & Grading Camera
A packing-line or handheld camera that grades fruit by ripeness and quality, sorting produce without human graders and keeping the line fast and independent through local grading. A vision model reads color, size, and blemish cues into a grade. Engineers learn visual quality assessment and mapping appearance to a commercial grade.
Precision Weed Sprayer
A camera on a sprayer or robot that distinguishes weeds from crop plants so herbicide is applied only where needed, cutting chemical use dramatically, with on-device speed essential to act while moving. A segmentation model separates crop rows from intruders in real time on an NPU board. The project teaches real-time segmentation and precision-agriculture actuation.
Microclimate Frost Predictor
A field station that learns the local microclimate and forecasts frost, heat, and humidity risk hours ahead, protecting sensitive crops with local prediction that beats coarse regional forecasts for a specific field. A compact forecaster projects conditions from on-site sensor history. Engineers learn short-horizon environmental forecasting on the edge.
Water Quality Sentinel
A floating or bankside node that monitors turbidity, pH, and dissolved oxygen and flags pollution or algal-bloom conditions early, with remote autonomous operation suited to rivers and reservoirs. A small model classifies water state and raises graded alerts. The project covers multi-parameter environmental sensing and early-warning classification.
Greenhouse Climate Optimizer
A controller that learns a greenhouse's dynamics and adjusts ventilation, shading, and irrigation to hold ideal growing conditions with minimal energy, running on-equipment so control is fast and resilient. A predictive model anticipates climate drift and acts ahead of it. Engineers learn predictive environmental control and multi-actuator coordination.
Forest-Fire Early-Warning Node
A solar node in wildland that fuses smoke, temperature, and gas sensing to detect fire ignition earlier than satellite or camera systems, with autonomous connected-when-possible operation suited to remote terrain. A small model separates a real ignition signature from cooking smoke and dust. The project teaches multi-sensor hazard detection with life-safety stakes.
Smart Home & On-Device Assistants
Consumer devices live in private spaces, so on-device AI is both a privacy promise and a reliability feature. The big 2026 shift is the local voice assistant: small language models such as Gemma now run on affordable edge boards like the Synaptics Coral Dev Board, so a home hub can understand and answer without ever sending speech to a server. These projects build that privacy-first ambient intelligence.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | On-Device Local Voice Assistant NEW | Synaptics Coral Dev Board + mic + speaker | Advanced | $200 |
| 02 | Presence-Aware Smart Lighting | ESP32-S3 + thermal/radar sensor | Intermediate | $45 |
| 03 | Appliance Energy Disaggregator | ESP32-S3 + current sensor | Advanced | $40 |
| 04 | Radar Gesture Light Switch | ESP32-S3 + 60 GHz radar sensor | Intermediate | $50 |
| 05 | Behavior-Learning Thermostat | ESP32-S3 + temp/humidity + display | Intermediate | $55 |
| 06 | Water-Leak & Pipe-Burst Detector | ESP32-S3 + flow + acoustic sensors + valve | Intermediate | $60 |
| 07 | Smart Smoke & Cooking Discriminator | ESP32-S3 + smoke/gas/temp sensors | Intermediate | $45 |
| 08 | Private Sound-Event Home Monitor | ESP32-S3 + MEMS mic | Beginner | $35 |
| 09 | Pet Recognition Door Controller | ESP32-P4 camera + servo | Intermediate | $60 |
| 10 | Air-Quality Predictor & Ventilator | ESP32-S3 + air-quality sensors + fan control | Intermediate | $50 |
On-Device Local Voice Assistant
A home hub that listens, transcribes, and answers questions or controls devices using a small language model running entirely on-device, so no conversation ever leaves the house, the defining privacy-first product pattern of 2026. Local speech recognition feeds an on-device SLM that interprets intent and generates a reply, with a compact tool layer for smart-home actions. Engineers learn to deploy a quantized language model on an edge NPU board and to wire speech, reasoning, and actuation into one local loop.
Presence-Aware Smart Lighting
A ceiling or corner sensor that infers real occupancy, and even how many people are present, to control lighting far more reliably than motion sensors that switch off when you sit still, with local inference that keeps lights instant and private. A low-resolution thermal or radar sensor feeds a small classifier robust to stillness. Engineers learn occupancy sensing beyond PIR and privacy-preserving presence detection.
Appliance Energy Disaggregator
A whole-home energy monitor that identifies which appliances are running from the aggregate power signature at the meter, no per-device sensors needed, with on-device disaggregation that keeps consumption data private. A model learns each appliance's electrical fingerprint and separates them from the combined signal. The project teaches non-intrusive load monitoring, a classic and hard signal-separation problem.
Radar Gesture Light Switch
A wall or lamp module controlled by simple hand gestures, wave to toggle, raise to dim, using a tiny 60 GHz radar sensor with no touch and no camera, so it works in the dark and preserves privacy. A gesture classifier maps motion to commands with strong false-trigger rejection, and local processing keeps it responsive. Engineers learn radar gesture recognition and designing an intuitive, low-error control surface.
Behavior-Learning Thermostat
A thermostat that learns household routines and preferences to preheat and precool automatically, saving energy without manual schedules, with on-device learning that keeps behavior data in the home. A compact model predicts occupancy and comfort demand from history. The project covers behavioral modeling and predictive comfort control.
Water-Leak & Pipe-Burst Detector
A node under a sink or by a water main that recognizes the acoustic and flow signature of leaks and bursts and shuts a valve before damage spreads, where local instant action is essential to prevent flooding. A small model separates a real leak from normal water use. Engineers learn acoustic and flow sensing fused for a fast protective response.
Smart Smoke & Cooking Discriminator
An alarm that tells dangerous smoke from harmless cooking steam and burnt-toast puffs, slashing the nuisance trips that make people disable detectors, with on-device decisions that keep the alarm fast and reliable. A model fuses particulate, gas, and temperature cues to judge true fire risk. The project teaches multi-sensor fusion for a life-safety device where false alarms have real cost.
Private Sound-Event Home Monitor
A hub that recognizes household sound events, doorbell, dog bark, running tap, breaking glass, and notifies without any always-on voice recording, where local processing is the privacy guarantee. A sound classifier covers a small event vocabulary while ignoring speech. Engineers learn multi-class sound-event detection and privacy-first product framing.
Pet Recognition Door Controller
A smart pet door and monitor that recognizes the household's pets and their activity, admitting them while excluding strays and logging behavior, with on-device recognition that keeps the door responsive and secure. A vision or collar-tag model identifies the specific animal. The project covers individual recognition and access control from a learned identity.
Air-Quality Predictor & Ventilator
A node that monitors indoor air, CO2, VOCs, particulates, and predicts when it will breach comfort thresholds, ventilating proactively, with local prediction and control that keep the home comfortable without cloud reliance. A small forecaster anticipates air-quality drift from occupancy and trends. Engineers learn environmental forecasting and closed-loop ventilation control.
Automotive & Mobility
Vehicles are safety-critical, real-time and often disconnected, the textbook case for edge AI, where a missed deadline is a safety event, not a dropped frame. In 2026 automotive-grade edge compute runs transformer-based perception for driver monitoring and road understanding on-device, and every inference must complete inside a hard, predictable window.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | Driver Drowsiness Detection | Jetson Orin Nano Super + IR camera | Advanced | $300 |
| 02 | Distraction & Phone-Use Monitor | Jetson Orin Nano Super + camera | Advanced | $300 |
| 03 | Forward Collision Warning | Jetson Orin Nano Super + camera | Advanced | $330 |
| 04 | VLM Road-Scene Understanding NEW | Jetson Orin Nano Super + camera | Advanced | $330 |
| 05 | Lane-Departure Warning | Jetson Orin Nano Super + camera | Advanced | $300 |
| 06 | Predictive Vehicle Diagnostics | ESP32-S3 + OBD-II interface | Intermediate | $45 |
| 07 | Tire & Road-Condition Sensor | STM32N6 + IMU + mic | Intermediate | $60 |
| 08 | EV Battery Health Estimator | STM32N6 + battery sensors | Advanced | $55 |
| 09 | Parking-Assist Object Detector | Jetson Orin Nano Super + camera | Advanced | $290 |
| 10 | Fleet Driving-Behavior Scorer | ESP32-S3 + IMU + GPS | Intermediate | $50 |
Driver Drowsiness Detection
An in-cabin camera that watches eye closure, blink rate, and head pose to detect drowsiness and alert the driver before microsleep causes a crash, with on-device inference mandatory because no cloud can meet the safety latency. A face and eye model runs continuously in the cabin's varied lighting. It is a flagship automotive-AI build and a lesson in robust facial analysis under harsh, changing light.
Distraction & Phone-Use Monitor
A camera that detects when a driver is looking away or using a phone and issues escalating warnings, complementing drowsiness monitoring while keeping cabin video private and responsive through local processing. A model reads gaze direction and hand-object interaction. Engineers learn gaze estimation and combining multiple cues into a graded attention score.
Forward Collision Warning
A dashcam-mounted system that detects vehicles and pedestrians ahead, estimates closing distance, and warns of imminent collision, running in a hard real-time loop where on-device speed is a safety requirement. Object detection and time-to-collision estimation run together on an NPU board. The project teaches real-time detection and closing-distance estimation on embedded compute.
VLM Road-Scene Understanding
A forward camera that describes the road scene in words, construction ahead on the right, cyclist entering the lane, using a small vision-language model to surface context a fixed detector would miss, all offline. This 2026-style build goes beyond boxes to language-level understanding on the edge. Engineers learn to deploy a compact VLM under automotive latency constraints and to turn scene descriptions into driver-facing alerts.
Lane-Departure Warning
A camera system that tracks lane markings and warns when the vehicle drifts without signaling, a core ADAS feature, with on-device processing that meets the tight control-timing budget. A model segments lane lines and geometry estimates lateral position. The project covers lane segmentation and turning geometry into a timely alert.
Predictive Vehicle Diagnostics
A dongle on the diagnostics port that learns a vehicle's normal signals and predicts developing faults, engine, transmission, emissions, before a warning light, giving early private warnings through on-device analysis. A model scores multivariate telemetry for anomalies. Engineers learn automotive time-series modeling and anomaly detection from bus data.
Tire & Road-Condition Sensor
A wheel or cabin sensor that infers road surface, dry, wet, ice, gravel, and tire condition from vibration and acoustics, informing traction and safety systems with local instant classification. A model maps the vibration signature to surface class. The project teaches surface classification from vibration and acoustic cues.
EV Battery Health Estimator
A battery-management add-on that estimates state of health and remaining range from charge-discharge patterns and temperature, improving on simple voltage lookups, with on-device estimation that keeps pack data local and the estimate live. A model tracks degradation trends over cycles. Engineers learn battery modeling and remaining-useful-life estimation.
Parking-Assist Object Detector
A rear or surround camera that detects obstacles, curbs, and pedestrians during parking and guides the driver, working reliably in tight low-speed scenarios where on-device speed is essential when maneuvering near people. A detector runs at low latency for close-range safety. The project covers close-range detection and driver guidance.
Fleet Driving-Behavior Scorer
A telematics device that scores driving behavior, harsh braking, cornering, acceleration, from motion sensors to coach fleet safety and cut insurance risk, with on-device scoring that keeps raw motion data local and sends only summaries. A model classifies driving events from IMU streams. Engineers learn event classification from vehicle motion and behavioral scoring.
Security, Safety & Surveillance
Security systems must keep working when the network is cut, often the attacker's first move, so local inference is a security property in itself. In 2026, NPU boards run person detection and video anomaly models on-device so alarms fire, faces are recognized and intrusions are caught without ever depending on, or exposing footage to, the cloud, and VLMs can now describe an event in plain language for faster triage.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | Smart Intrusion Detection Camera | Raspberry Pi 5 + AI Kit + IP camera | Advanced | $190 |
| 02 | Access-Control Face Recognizer | Jetson Orin Nano Super + camera | Advanced | $290 |
| 03 | VLM Perimeter Anomaly Describer NEW | Jetson Orin Nano Super + camera | Advanced | $310 |
| 04 | Weapon & Threat-Object Screener | Jetson Orin Nano Super + camera | Advanced | $320 |
| 05 | PPE Compliance Monitor | Raspberry Pi 5 + AI Kit + camera | Advanced | $190 |
| 06 | Vibration-Based Break-In Sensor | Nordic nRF54L15 + accelerometer | Intermediate | $35 |
| 07 | Aggression & Fall Detector for Care | Jetson Orin Nano Super + camera | Advanced | $290 |
| 08 | Acoustic Gunshot Locator | STM32N6 + mic array | Advanced | $90 |
| 09 | Tailgating & Occupancy Sensor | ESP32-S3 + ToF/thermal sensor | Intermediate | $60 |
| 10 | License-Plate Watchlist Alerter | Jetson Orin Nano Super + camera | Advanced | $320 |
Smart Intrusion Detection Camera
A security camera that distinguishes people from animals, blowing leaves, and headlights, cutting the false alarms that plague motion-only systems, recording only meaningful events and working offline. On-device detection classifies the trigger before alerting, keeping footage private. It is a core surveillance-AI build and teaches robust person detection amid outdoor clutter.
Access-Control Face Recognizer
A door unit that recognizes enrolled faces to grant entry, storing face templates locally so no biometric data ever reaches a server, which makes local matching both a privacy and a security guarantee. A detector locates the face and an embedding model matches it against an on-device gallery. Engineers learn face embeddings and secure on-device biometric matching.
VLM Perimeter Anomaly Describer
A fixed camera watching a fence or yard that learns the normal scene, flags anomalies, and uses a small vision-language model to describe what it sees, person climbing the north fence, bag left by the gate, for faster human triage, all on-device. Combining anomaly scoring with language output is a distinctly 2026 capability. The project teaches video anomaly detection plus on-edge VLM captioning for security.
Weapon & Threat-Object Screener
A checkpoint camera that flags visible threat objects to alert security staff, adding an automated layer to manual screening with local inference for immediate private screening. A detector localizes target objects and raises graded alerts for human confirmation. Engineers learn object detection for safety and human-in-the-loop alerting.
PPE Compliance Monitor
A worksite camera that verifies workers wear required protective equipment, hard hats, vests, goggles, and logs compliance for safety enforcement, with on-device processing that keeps worker footage local. A detector checks for each item per person. The project covers multi-attribute detection and turning it into a compliance record.
Vibration-Based Break-In Sensor
A discreet sensor on a door, window, or safe that recognizes the vibration signature of forced entry, prying, drilling, glass impact, while ignoring normal knocks and slams, alerting instantly and locally even if lines are cut. A model classifies the vibration event. Engineers learn vibration event classification for security.
Aggression & Fall Detector for Care
A ceiling sensor in care settings that detects falls and aggressive incidents from motion and pose, alerting staff while protecting resident dignity with on-device, non-recording analysis. A pose model recognizes the target events, and local processing is the privacy foundation for a sensitive environment. The project teaches pose-based event detection with strong privacy constraints.
Acoustic Gunshot Locator
A network of nodes that detect gunshots and estimate direction from the acoustic wavefront, alerting responders in public-safety deployments with on-device detection that gives the speed such alerts demand. Each node classifies the impulsive event and shares timing for localization. Engineers learn impulsive-event detection and multi-node acoustic localization.
Tailgating & Occupancy Sensor
A doorway sensor that counts people entering and detects tailgating, more people passing than badges scanned, for secure-facility access control, with local processing that keeps entry data private and instant. A small model counts and tracks crossings. The project covers people counting and rule-based security logic on the edge.
License-Plate Watchlist Alerter
A camera that reads plates and checks them against a locally stored watchlist, alerting on matches without sending any footage off-site, so both the watchlist and the footage stay private. Detection and OCR feed an on-device lookup. Engineers learn plate recognition and secure on-device list matching.
On-Device Generative AI & Edge MLOps
The biggest change of 2026: generative models now run on the edge. Small language models like Gemma, compact speech recognizers, and tiny multimodal transformers fit on affordable NPU boards, enabling local chat, retrieval and transcription with no cloud. This family teaches the frontier skills, running and fine-tuning SLMs on-device, edge RAG, agentic tool use, and the MLOps to ship and update it all, alongside the TinyML fundamentals everything rests on.
| # | Project | Core hardware | Level | Cost |
|---|---|---|---|---|
| 01 | On-Device SLM Chat Assistant NEW | Synaptics Coral Dev Board | Advanced | $130 |
| 02 | Edge RAG Document Assistant NEW | Raspberry Pi 5 + AI Kit | Advanced | $160 |
| 03 | Offline Whisper-Class Transcriber NEW | Raspberry Pi 5 + AI Kit + mic | Advanced | $160 |
| 04 | Agentic Edge Tool-Caller NEW | Synaptics Coral Dev Board + peripherals | Advanced | $170 |
| 05 | On-Device Personalization with LoRA NEW | Raspberry Pi 5 + AI Kit | Advanced | $155 |
| 06 | MNIST Classifier on Bare Metal | Arduino Nano ESP32 | Beginner | $30 |
| 07 | Sensor-Data Logger & Dataset Builder | Arduino Nano ESP32 | Beginner | $35 |
| 08 | Quantization & Latency Benchmark Rig | STM32N6 Nucleo board | Intermediate | $60 |
| 09 | Reusable Anomaly-Detection Node | STM32N6 + IMU | Intermediate | $55 |
| 10 | End-to-End Edge MLOps Pipeline | Raspberry Pi 5 + cloud tooling | Advanced | $90 |
On-Device SLM Chat Assistant
A self-contained assistant that answers questions and holds a conversation using a small language model running entirely on an edge board, proving that useful generative AI no longer needs the cloud. You deploy a quantized SLM such as a Gemma-class model, manage its context window and memory footprint, and stream tokens locally. It is the hello-world of edge generative AI and teaches quantization of language models, prompt handling, and the real latency and memory trade-offs of on-device inference.
Edge RAG Document Assistant
A private assistant that answers questions about your own documents by retrieving relevant passages and feeding them to a local small language model, so confidential files are never uploaded anywhere. You build an on-device embedding index, a retrieval step, and a generation step, the full retrieval-augmented-generation loop on the edge. Engineers learn embeddings, vector search on constrained hardware, and grounding an SLM's answers in local data.
Offline Whisper-Class Transcriber
A device that turns speech into accurate text fully offline using a compact automatic-speech-recognition model, useful anywhere audio is sensitive or connectivity is absent. You deploy a quantized sequence model, manage streaming audio buffers, and balance vocabulary and accuracy against flash and RAM. It is a focused lesson in running a transformer speech model on the edge and the engineering of real-time on-device transcription.
Agentic Edge Tool-Caller
A local assistant that not only answers but acts, checking a sensor, toggling a relay, querying a local database, by having a small language model choose and call tools, the on-device version of the agentic pattern sweeping AI in 2026. You define a tool interface, constrain the model's outputs, and safely execute its chosen actions. Engineers learn structured generation, tool-use orchestration, and guardrails for an autonomous local agent.
On-Device Personalization with LoRA
A device that adapts a pretrained model to its user or environment from a handful of local examples using lightweight fine-tuning, without ever sending data to the cloud. You attach a small trainable adapter and update only that, demonstrating personalization at the edge. Engineers learn parameter-efficient fine-tuning and lightweight on-device training, a genuine frontier skill for edge AI.
MNIST Classifier on Bare Metal
The classic first deployment: train a small digit classifier, quantize it to INT8, and run it on a microcontroller with no operating system, proving the full train-to-device pipeline end to end. The value is entirely in the process, exporting, quantizing, fitting the tensor arena, and measuring latency, and it remains the fastest way to internalize the deployment toolchain before tackling real sensors.
Sensor-Data Logger & Dataset Builder
A device that captures and labels sensor data, motion, audio, environmental, to build the custom datasets every real project needs, since public data rarely matches your exact sensor and mounting. It streams labeled samples for training and supports on-device replay. Good data beats a clever model, and this teaches the underrated discipline of collecting it well, from capture to labeling workflow.
Quantization & Latency Benchmark Rig
A test harness that runs the same model at different precisions, float, INT8, and lower, and measures accuracy, latency, and memory on real hardware, making the deployment trade-offs concrete and repeatable. It automates the compare-and-report loop across model variants and surfaces per-layer timing. Understanding these trade-offs is the heart of edge ML, and this project teaches quantization effects and rigorous on-device benchmarking.
Reusable Anomaly-Detection Node
A general-purpose node that learns the normal pattern of any attached sensor and flags deviations, a reusable template behind countless predictive-maintenance and monitoring builds. It uses a small autoencoder-style model that needs only healthy data, so mastering this one pattern unlocks a whole family of applications. Engineers learn one-class anomaly detection as a transferable building block.
End-to-End Edge MLOps Pipeline
A complete workflow that versions data and models, trains, quantizes, tests on hardware, and deploys over the air, the production backbone that turns a demo into a maintainable fleet, now including delivery of generative models to the edge. It automates the path from commit to device. Shipping and updating edge AI at scale is its own engineering discipline, and this project teaches edge MLOps and over-the-air model delivery.
Choosing your first build
If you are new to embedded AI, start with the foundations in the final family, deploy a digit classifier to bare metal, then build a custom wake-word engine. Those two teach the entire toolchain: capturing data, quantizing to INT8, fitting a tensor arena, and measuring latency and power on real silicon. Everything else, including the new on-device generative and physical-AI builds, is a variation on that loop with a different sensor, a bigger model and a different decision.
From there, follow the constraint that interests you. Care about privacy and battery life? The wearables and audio families reward it. Want to see AI move something in the physical world? Robotics and the actuating agriculture projects close the loop from perception to motion. Chasing clear return on investment? Predictive maintenance turns a few dollars of MEMS sensor into avoided downtime that dwarfs the hardware cost.
