KreditBee — Whisper + LLAMA3 audit

How to improve

• →17% of calls have one-sided audio — verify telephony stereo recording config (Exotel/Knowlarity often mixes both legs on a single channel).
• →10% of calls are ≤5s — skip these from the LLM pipeline entirely to cut cost and noise.
• →Whisper struggles with code-mixed Hindi-English; switch to Whisper-large-v3 + `language='hi'` hint, or use Sarvam/Reverie for Indian-language ASR with diarization built in.
• →Add VAD pre-filter (silero-vad) — drops empty calls before they hit the GPU queue.

How to improve

• →LLAMA3 marks 91% of calls 'incorrect' — base rate is implausible. The prompt likely lacks a clear definition of each disposition; supply the KreditBee SOP verbatim as a system message.
• →`correct_status` returns free-form strings like '-', '-interested-', '-not_interested-'. Constrain with an enum + JSON-schema-guided decoding (llama.cpp grammar / Outlines).
• →Force structured output and reject on parse failure — 1 row already returned `+++ {...}` prefix which would break a downstream ETL.
• →Add a 'min_evidence' rule: if total transcript < 15 words, return `insufficient_evidence` instead of judging.
• →Switch to LLAMA3.1-8B-Instruct or Mixtral with explicit chain-of-thought + final JSON — base LLAMA3 hallucinates dispositions on Hindi transcripts.