AI that runs where your operations happen

Camera-based cycle counting

The problem

Manual cycle counts are slow, error-prone and pull staff off picking. Mis-counts become stockouts and write-offs.

What we build

A cart- or shelf-mounted Jetson camera counts items per bin, reads bin labels and flags variances against the WMS before syncing counts to the ERP.

Measured by

Count time per bin, accuracy against manual count, staff hours returned to picking

• Jetson Orin Nano
• Sony IMX477 camera
• YOLO + TensorRT
• Label OCR
• REST / MQTT

Low-cost stock-level monitoring

The problem

Critical parts run out between counts, and fixed cameras are too expensive to put on every shelf.

What we build

Battery-powered ESP32 nodes pair a load cell with a small camera, estimate fill level on the device and trigger replenishment below a threshold. Images are captured only on change.

Measured by

Battery life per node, cost per bin, stockouts avoided

• ESP32-S3
• OV2640 camera
• HX711 load cell
• On-device TinyML
• MQTT
• Deep sleep

Wearable hands-free picking

The problem

Handheld scanners slow pickers down, and juggling a device with stock causes errors.

What we build

A clip-on ESP32 device reads out the next pick and listens for spoken confirmations. Wake word and command recognition run on the device, so it works on patchy warehouse Wi-Fi.

Measured by

Picks per hour, pick error rate, battery life per shift

• ESP32-S3
• I2S MEMS microphone
• On-device keyword spotting
• BLE / Wi-Fi to WMS

Vision-guided pick-and-sort robot

The problem

Sorting mixed small parts is repetitive and hard to staff, and industrial robot cells cost too much for a pilot.

What we build

A Jetson detects parts and plans each pick. An ESP32 co-processor drives the servos with calibrated limits and a hardware watchdog, so the arm stops safely if the main computer stalls.

Measured by

Picks per minute, pick accuracy, build cost against an industrial cell

• Jetson Orin Nano
• ESP32 co-processor
• PWM servo driver
• Hardware E-stop
• Python robot API

Automated receiving at the dock door

The problem

Receiving staff scan labels one at a time, and damaged cartons are often found days later when claims are harder to make.

What we build

A camera over the dock door reads labels and barcodes, matches them to the expected ASN and flags crushed or torn cartons with a timestamped photo.

Measured by

Receiving time per pallet, damage claims recovered, aSN mismatches caught

• Jetson Orin Nano
• Industrial camera + strobe
• Barcode and OCR models
• Damage detection
• ERP API

Conveyor predictive maintenance

The problem

A conveyor breakdown halts fulfilment, and failing motors give warning in sound and vibration that nobody is listening for.

What we build

ESP32 sensor nodes clamp onto motors and rollers, learn each machine's normal pattern and raise an alert only when it changes. Raw audio never leaves the node.

Measured by

Failures caught early, unplanned downtime, cost per sensor node

• ESP32-S3
• MEMS accelerometer + mic
• On-device FFT + anomaly model
• MQTT
• Teams / email alerts

Fleet platform for edge AI devices

The problem

Edge pilots fail at scale: devices drift, updates break drivers and a dead board takes days to rebuild.

What we build

Every device is built from a pinned golden image, reports its health at boot, receives signed over-the-air updates and can be re-imaged from documented scripts.

Measured by

Devices under management, rebuild time, update success rate

• JetPack image pipeline
• Docker
• systemd watchdogs
• ESP-IDF secure OTA
• Prometheus + Grafana