Meeting Notes Agent

What it does

• Ingests meeting transcripts (or audio via STT) and produces structured summaries
• Outputs agenda items, action items, decisions, and key quotes in canonical format
• Validates all outputs via Zod before they reach consumers
• Renders deterministically for consistent downstream use

Architecture

┌──────────────────┐     ┌───────────────────┐     ┌──────────────────┐
│  Transcript /    │────►│  Extraction       │────►│  Schema          │
│  Audio (STT)     │     │  Pipeline (LLM)   │     │  Validation      │
└──────────────────┘     └───────────────────┘     │  (Zod)           │
         │                         │               └─────────┬────────┘
         │                         │                         │
         ▼                         ▼                         ▼
┌──────────────────┐     ┌───────────────────┐    ┌──────────────────┐
│  Trace /         │     │  Deterministic    │    │  JSON / Web      │
│  Audit Log       │     │  Renderer         │    │  Output          │
└──────────────────┘     └─────────┬─────────┘    └──────────────────┘
                                  │
                                  ▼

Reliability & Guardrails

• Zod schema validation catches malformed LLM outputs before they reach consumers
• Eval harness runs golden-case and invariant checks (e.g., all action items have owners, no empty sections) on every change
• Trace artifacts support audit and debugging

Key Engineering Decisions

• Zod for schema validation instead of manual parsing — Catches malformed LLM outputs at parse time; single source of truth for types and runtime checks.
• Deterministic renderer before JSON output — Downstream consumers get stable formatting; avoids drift from prompt changes or model variance.
• Trace artifacts stored per extraction run — Audit trail for compliance; enables debugging without re-running the LLM.

Failure Modes Handled

• LLM returns invalid JSON or missing required fields — caught by Zod parse
• Hallucinated action-item owners — invariant check fails in eval harness
• Empty sections when transcript is sparse — default fallbacks + invariant for non-empty
• Model downgrade introduces new failure modes — golden cases catch regressions

How to demo in 2 minutes

1. Clone the repo and install deps: npm install
2. Set OPENAI_API_KEY in .env
3. Run the dev server: npm run dev
4. POST a sample transcript to /api/extract or use the web UI
5. Inspect the structured output and trace artifacts in the response

Links