03 How it works
Built to answer commercial, sourcing, and pricing questions in 3 seconds, and keep itself current without manual wiki chores.
The operating loop
When
A team member asks Copilot, Gemini, ChatGPT, Claude, or /app
Then
Hooks inject matching prices, Living Briefs, open conflicts, and Do-Not-Contact alerts in ~2.4s before the AI writes a word.
Staff get verified numbers with source citations instead of guessing.
When
A decision lands via WhatsApp, WeChat/WeCom, Telegram, or Email CC
Then
SQL narrowing checks existing Living Briefs for conflicts and sends a private 1-to-1 card (Approve, Edit, or Ignore), never replying-all on external threads.
Casual hypotheticals are filtered out automatically.
When
A draft quote runs pre-send or morning markets move
Then
Quote Shield flags FX, margin-floor, DNC, and OEM leaks in 5 seconds; 08:00 triggers watch EUR/CNY, USD/CNY, LME copper, and tariffs against contract clauses.
People still tap Approve before anything edits canonical memory.
Capture circuit
WhatsApp voice notes, WeChat QC photos, Telegram sheets, and Email CCs run through SQL narrowing and a 1-to-1 owner approval tap before touching Postgres.
Voice note (EN/FR/繁/简): "Hold Dongguan copper surcharge at 3% for Q1 Rexel quotes."
Conflict check: AO2027 Offer v3 currently states 2%. Stage update to 3%?
Eight architecture layers
Every layer sits inside your own Postgres project and works identically regardless of which permitted AI answers the prompt.
Lockout drill
Every quarter we unplug your main AI, connect your fallback AI over MCP, run 10 real company questions, and verify identical answers and rule refusals in under 30 minutes.
Next