Conversation Intelligence
GPT-5 + pgvectorAn AI system I designed and built that turns unstructured conversations into structured, coachable insight: transcript ingestion, strict-schema LLM analysis, embedded retrieval, and a role-aware coaching chat grounded in citations. I own the architecture, the data model, the pipeline, the backend, and the UI.
The Problem
Conversations hold the signal. Nobody has time to mine it.
Sales and support calls are full of objections, missed questions, and language patterns that decide outcomes. Reviewing them by hand does not scale, and generic transcription tools stop at words on a page.
The hard part is not transcription. It is turning a raw transcript into trustworthy, structured metrics and grounded feedback that a person will actually act on.
The System
Transcript in. Structured coaching out, in under a minute.
A transcript is ingested atomically, analyzed by GPT-5 against a strict JSON schema, embedded into a pgvector store over a domain knowledge corpus, and served back through a role-aware chat that answers with token-highlighted citations. Metrics like talk ratio and open-ended-question rate land on every conversation automatically.
How it works.
Ingest
Atomic transactional ingestion via Postgres RPCs. No orphan records under concurrent writes.
Analyze
GPT-5 with strict JSON-schema output. Talk ratio, question rate, objection handling, language patterns.
Retrieve
pgvector embedded retrieval over a domain corpus, with token-overlap citation scoring.
Coach
Role-aware chat that answers grounded in the transcript and corpus, with highlighted citations.
Tech stack.
AI & Retrieval
- GPT-5 structured output
- OpenAI embeddings
- pgvector
- Citation scoring
Backend
- PostgreSQL
- Postgres RPCs
- Row-level security
- Signed-URL uploads
Frontend
- Next.js
- React
- TypeScript
- Tailwind CSS
Async & Infra
- Inngest step functions
- Audio pipeline
- Structured logging
- Vercel