Adam Abu-Taqa
Current hardware: Geniatech XPI-3566-ZERO — Rockchip RK3566 (4× Cortex-A55 @ 1.8 GHz), 8 GB single-channel LPDDR4, Debian 11. Chosen for cost; benchmarking on it produced the first public llama.cpp numbers for this chip (log #1).
Where testing has landed (so far): a fine-tuned Llama-3.2-3B is the model this product needs — stock models at any tested size fail tool-calling, and fine-tuning moved it from 5/13 to 10/13 with zero fabrications (log #5). Measured on the RK3566 the 3B runs 2.62 tok/s, below the 3.5 tok/s conversational kill line, so the compute is moving to RK3588-class silicon. That conclusion is a measurement, not a guess, and the logs show the path to it.
Architecture: whisper.cpp for speech in · Llama-3.2 (fine-tuned) for language · GBNF grammar-constrained decoding so tool outputs are structurally valid · routing done in ordinary code, not by the model · safety-critical reference content retrieved and read verbatim, never paraphrased · a code-side gate refuses out-of-corpus questions without ever invoking the model.
Status: pre-demo. Schematic complete (five sheets, KiCad), PCB not started, audio path (I2S) is the current bring-up blocker.
