Browser VAD -> Yandex SpeechKit ASR -> command classifier -> Qwen LLM -> JSON Schema validation -> REST API write to MIS
Built a hands-free clinical voice assistant integrated with the STOMMIS dental MIS, piloted at SPb GBUZ SP No. 8 with a signed institutional adoption agreement.
A TF-IDF + logistic regression command classifier trained on 8,000 synthetic transcriptions gates LLM invocation, reducing ASR and LLM cost by 95-97%.
Qwen output is constrained to a strict intent + slots JSON contract, validated with JSON Schema and an intent whitelist before any MIS write. No medical record is modified unless the full pipeline validates cleanly. End-to-end command latency is about 2-3 seconds.