Read everything, store almost nothing
Raw conversation text expires after processing; only abstract summaries, entities, relationships, confidence and provenance survive.
ICB Radar
A privacy-conscious WhatsApp intelligence pipeline that turns selected coffee-community conversations into alerts, signals and an evergreen knowledge wiki.
system · active
Product owner and builder · Updated Aug 2026
- Conversation intelligence
- Privacy
- Knowledge graph
- Local AI
01 · Problem
Specialty-coffee groups contain valuable launches, questions and emerging needs, but manually reading every message is expensive and permanently archiving private conversation would be irresponsible.
02 · System
An allowlisted connector passes short-lived text into a conversation pipeline that resolves entities, extracts privacy-safe signals and writes both a personal intelligence feed and an evergreen Markdown wiki.
03 · Outcome
The architecture enforces ephemeral raw text, metadata-only logs, derived long-term stores, a review dashboard and human approval before any external action.
Context
The system treats privacy as an architectural property. Exact messages, participant names, phone numbers and media are excluded from permanent storage by design.
Architecture
- 01Allowlisted WhatsApp connector
- 02Redis-backed ephemeral conversation buffer with TTL
- 03Entity resolver and signal engine
- 04Postgres context graph and privacy-safe derived records
- 05Markdown wiki writer, scheduler and private review dashboard
Decisions & trade-offs
Evidence
- Raw-payload fields are rejected from operational logs
- Database migrations contain no raw-message columns
- Unit tests run without a live WhatsApp account
- Outbound automation is restricted to GT's private intelligence feed
What this taught me
- Privacy constraints can improve the information model by forcing a distinction between evidence, claim and durable insight.
Follow the system
Connected work is resolved from the shared content graph.