I build at the intersection of embedded systems and AI - firmware, IoT hardware, and developer tools that make on-device ML less painful.
Check if an AI model fits on a microcontroller - before you flash it. Instant fit verdicts across 31 boards, byte-exact memory measurement using the real TFLite Micro runtime (native and WebAssembly), quantization previews, and a CI action.
🌐 Try it in your browser · pip install mcufit
I work on speech recognition and tool-calling models that run on embedded Linux boards with under 200 MB of RAM, and no cloud round trip. Mostly the unglamorous half of it: what actually fits in memory, what real-time factor an ARMv7 core can hold, and when the honest answer is that the model does not fit.
What TensorFlow Lite Micro actually costs on real hardware. The ground truth mcufit is checked against: 54 measurements across 4 models and 3 boards, per model and per layer. Anyone can add a board.
Format-aware compression without the compression expertise.
An agent reads your data, writes the OpenZL data description, trains a compressor, verifies the round-trip byte-for-byte, and reports the savings against zstd and gzip. 50% off a CSV that zstd -3 had already compressed.
- loan-emi-negotiator - an agent that drafts a lender-specific rate-negotiation email
- docpilot - document collection loop for vehicle-loan officers
- IoT & embedded builds - laser turret, fire-alert Telegram bot