ML System Design at Chinor

Выводы и как готовиться

• The case is a compact ASR-to-action ML System Design: audio is not the final материал, the booking table update is.
• The strongest design separates accept/reject routing, ASR quality, entity extraction and slot-write validation instead of hiding everything behind one model.
• Low-resource languages change the problem: in-domain transcription data and label quality become core system components.
• Production concerns include slot conflicts, idempotent writes, large audio transfer, queue processing and per-stage inference cost.