01.01
On-Device Submission Intake Console
Show-floor demo · local document AI · Aug — Oct 2026
- Built a full-screen intake console that turns a dropped broker packet — PDFs, scans, emails — into a structured submission record without sending a byte off the machine.
- Designed a deterministic-first pipeline: a Python sidecar reads text geometry, tables, and checkbox ink; a local small language model fills each field against a per-document-type schema at temperature 0, so the same page yields the same JSON.
- Implemented cross-document reconciliation that turns conflicts, corroborations, and unreadable fields into review items, with click-to-source that jumps to the exact bounding box a value came from.
- Wrote a bridge that fronts Foundry Local and llama.cpp/Ollama behind one OpenAI-compatible endpoint so the pipeline could use the GPU on an ARM box with no CUDA provider — a full packet went from ~10 minutes of model time on CPU to about a minute.
- Instrumented the browser's fetch for a wire log that proves nothing left the machine, plus live network state, a $0.00 metered-cost counter, and a Postgres case store with seen-before cross-checks per insured.
- Added presenter controls for a live stage run: a frozen demo packet, a golden fallback, a pipeline inspector showing every prompt and raw response, a client list grouped by insured, and JSON export, with unit tests over the pipeline logic.