⚡ Flash — your meeting's memory, as a living canvas

Flash joins your Google Meet as a third participant. It listens passively (and watches shared screens), turning the conversation into a live graph on a web canvas. Say "Hey Flash" and it wakes up, answers with voice, and draws diagrams grounded in the live discussion + prep docs. When the meeting ends it writes the summary and files it into your dashboard history — one context, reused live and after.

Built with Superlinked (semantic retrieval), n8n + Gemini (the brain), and SLNG (voice).

Architecture

Dashboard (Face) ──"Add Flash"──▶ P1 /join ──Playwright──▶ Google Meet
  P1 ears (SLNG STT) ─▶ POST /ingest ─▶ P2 (SQLite + Superlinked)   [memory]
                      └▶ POST /events ─▶ Face canvas                 [live nodes]
  "Hey Flash …" ─▶ P1 ─▶ n8n /agent ─GET /retrieve─▶ P2 ─Gemini─▶ {text, mermaid, sources}
                          ├─ SLNG TTS speaks it into the call
                          └─ Face renders the diagram node live
  End meeting ─▶ Face /summarize ─▶ n8n /finalize ─▶ summary/decisions/actions
              └▶ Face /sync-memory ─▶ P2 /sources (type:canvas) ─▶ Superlinked  [query the graph]
  Post-meeting Q&A ─▶ n8n /ask ─▶ P2 /retrieve (incl. source:canvas) ─Gemini─▶ grounded answer

Components: P1 agent/ (Meet bot), P2 src/ (retrieval service), P3 Flash-Agent.json (n8n workflow), P4 face/ (Next.js canvas). Contracts: shared/contracts.ts + ARCHITECTURE.md.

Setup

cp .env.example .env          # fill GEMINI_API_KEY, SUPERLINKED_API_KEY, SLNG_API_KEY
npm install                   # P2 + n8n deps (repo root)
( cd agent && npm install && npx playwright install chromium )
( cd face  && npm install && cp .env.example .env.local )

Run (4 terminals)

#	Service	Command	Port
1	P2 Retrieval	npm run seed:retrieval && npm run dev	3000
2	P3 Brain (n8n)	npm run n8n:import && npm run n8n:start	5678
3	P4 Face	cd face && npm run dev -- -p 3001	3001
4	P1 Agent	cd agent && npm start	8001

After the first n8n:import, activate the workflow once so its production webhooks register: N8N_USER_FOLDER=.n8n-local ./node_modules/.bin/n8n update:workflow --id=flash-agent-p3 --active=true then restart n8n (terminal 2). All five webhooks then live at http://localhost:5678/webhook/{agent,ask,finalize,events,vision}.

Smoke tests

curl "localhost:3000/retrieve?meetingId=m_demo&query=budget&k=3&mode=auto"   # superlinked-rerank
curl localhost:3000/meetings                                                  # history list
curl -XPOST localhost:5678/webhook/agent -H 'content-type: application/json' \
  -d '{"meetingId":"m_demo","requestText":"make a diagram of the budget"}'    # grounded mermaid
curl localhost:3001/api/meetings                                              # dashboard data

Demo (the money shot)

1. Open http://localhost:3001, enter a Meeting ID + the Google Meet link, click Add Flash.
2. Two people talk; utterance nodes appear live on /m/<id> as they speak.
3. Say "Hey Flash, make a diagram of the budget" → Flash speaks the answer and a diagram node appears on the canvas, grounded in the conversation (cites speakers).
4. Ask "Hey Flash, what did we decide?" → grounded spoken answer.
5. Click End meeting → summary / decisions / action-item nodes; the meeting lands in the dashboard's Recent meetings. Reopen it later — the graph is restored and re-queryable.

Notes

• Single root .env feeds P1, P2 and n8n; face/.env.local feeds the Next app. Secrets never commit.
• Grounded, not hallucinated: every live answer/diagram retrieves from Superlinked first and cites sources.