Pulse — Demo Day Command Center
Everything from walking on stage to Q&A. One tab. AABW 2026 · freeze Jul 12, 09:00 ICT.
0 · Pre-flight (10 min before)
- Terminal in
C:\nikhil\code\prodigi\pulse\app → npm run dev (app on localhost:3000)
- Fresh state:
npm run demo:reset — keeps enrichment, never re-run enrichment itself
- Open these tabs in order: this page · app · deck · Langfuse (your bookmark — URL in
app\.env) · Canva folder
- Budget check: full rehearsal ≈ $0.24 → rehearse freely
1 · Open with the deck
2 · Module 1 — Consumer Intelligence (90s)
1Overview"796 customer signals in one stream, AI-tagged — 592 real reviews scraped from Hasaki and Watsons, plus Guardian site reviews and CS transcripts built to spec, since their internal data isn't public."
2Point at health + "What changed this week""Nobody reads 800 comments. Pulse tells you what moved and why."
3Click Scan for issues — steps stream live"Sentinel agent: detects the spike in code, cites actual customers, routes to the right team. In production this runs on a schedule — the button is so you can watch it."
4Alerts — open the new one
"Delivery complaints 3× normal. Root cause, cited quotes, routed to Ops. Humans decide — AI drafts."
5Benchmarks"Same pipeline pointed at Hasaki and Watsons — in their customers' own words."
6AI Activity →
Generate exec summary"The monthly reporting deck in one click — the 70–80% manual-effort cut."
3 · Module 2 — Creative Intelligence (90s)
1Performance"$2.1M spend, 60 creatives, 8 themes. Two numbers per theme: week-one return and lifetime return."
2Dual badges → event-offer card"Week-one ROAS understates true value ~2.7×. Event-offer looks weak at 1.11 — but it's all whales. Installs versus valuable players."
3Ad Library (new)"21 real ads scraped from Meta's Ad Library — this is the live market we're generating against."
4Click Scan for insights, open fatigue insight"Flags computed in code, AI writes the explanation. Feature-list burned ~$48.9k in 30 days — recommendation in pLTV-weighted ROAS, the brief's own win condition."
5Generate variants on the winner"Briefs + copy per channel and market, graded SHIP/REVISE/KILL before a dong is spent. ~72 seconds — time-to-first-variant is the win condition."
7Reviewer
Approve →
Design A/B test →
A/B Tests →
Read out ×3
"Sample size from a power calculation — the LLM never touches the math. Ship / kill / iterate: 3 of 3 closed with a statistically-sound decision."
8Audiences (sample — say "this is the identity spine we'd build")"Match the Facebook click to the in-game player and you price campaigns in lifetime value, not installs — TikTok's cheap installs, worst pLTV:CAC."
4 · Observability proof (when judges ask "how do you trust it?")
- AI Activity (Consumer) / AI Activity (Creative) — every run: steps, tokens, cost
- Click any trace link on a run row → opens Langfuse — exact prompt, answer, dollars. Exec-summary runs on LangChain.
- Line: "Numbers come from SQL, the LLM only explains. Every agent run is traced — here's one."
5 · Q&A ammo
796 real signals · 4 source types
Eval gate: 96% sentiment / 82% topic
592 scraped competitor reviews
21,600 ad-metric rows
Portfolio 1.42 ROAS / 4.04 LTV-ROAS
Budget moves: +0.45 LTV-ROAS, ≤30% daily cap
z-tests in code: ship / kill / iterate, 3/3 sound
Variants: ~72s, length-linted, VN+EN
Demo rehearsal ≈ $0.24
White-label: brand only in Workspace config
Math answers: sample size = two-proportion power calc (α .05, power .8); readout = two-proportion z-test; verdict decided in code, Sonnet narrates.
6 · If something breaks
- App wedged → kill node on :3000,
npm run dev again, npm run demo:reset
- Sentinel/insights already ran → that's fine, open the existing alert/insight instead of re-running
- Deck offline → decks are also in repo:
pulse\docs\deck-hackathon.html (open the file directly)
- Full coaching:
pulse\docs\READING-THE-DATA.md · LangChain tour: docs\LANGCHAIN.md · budget: docs\COSTS.md
Pulse · one platform, five briefs — LISTEN · WATCH · COMPARE / LEARN · MAKE · PROVE