ReportAI
An agent that turns a voice note into the finished corporate report already filled into each client's own Word template — meeting minutes, a visit report — and delivers it on the chat channel they already use, no new app to install.
The Problem:
Most AI meeting-note agents assume the meeting is a scheduled video call with a bot sitting inside it. That structurally misses the most common case: an in-person visit or a phone call, recapped afterwards from a phone, with no laptop and no meeting link. And almost none of them let a company upload its own Word template — they return their own generic summary instead of the actual document the company already uses.
- Existing meeting-note agents require a bot inside the video call — useless for in-person visits or phone calls
- No tool fills a company's own Word template; they all return a generic summary format instead
- Documenting a visit still meant transcribing by hand and formatting the report afterwards
Objective:
Build an agent that runs on the chat platform each team already has open, with no new app or integration, and that's faithful to each client's real document — not a generic summary:
- Each client defines their own template and document types from an admin panel — no code change to onboard
- Structured extraction with schema validation, not just generated text: a wrong date or a dropped agreement in a formal report is a real failure
- Mandatory human review before sending, approved or corrected by replying to the message itself
Stack & highlights:
- Backend: FastAPI (Python)
- Agent orchestration: LangGraph
- Extraction: Claude — structured output / tool-calling with schema validation
- Speech: Groq Whisper
- Database: Self-hosted PostgreSQL
- Documents: docxtpl + self-hosted Gotenberg — Word template → final document
- Admin panel: Vue/Nuxt · Vuetify · Tailwind
- Infra: Docker Compose · Traefik
- Full chain: transcribe → extract with schema validation → fill the client's own template → deliver on the same channel
Outcome:
A field rep finishes a visit and records a 90-second voice note from the car. Minutes later, the report lands in the right inbox — filled into the company's exact template, not a generic AI summary:
- In production with 3 active clients, processing 10 to 15 reports a month
- Faithful to the real document: extraction is validated against each template's schema before generating it — a wrong date or a dropped agreement doesn't slip through
- Genuinely multi-tenant architecture: onboarding a new client means configuring a template, document types, and a channel from the admin panel — no code touched
- Own eval suite guarding extraction quality on every change
- Live public demo at reportai.vazquezlabs.com
