🦈 Shark Room

🏆 1st Place — Cursor Community Hackathon @ SAIT (May 2026) Built by Felix Gabriel Montañez & Xujun Wang

Originally developed at felix101930/CursorHack; this repo is a maintained mirror.

Every AI pitch tool tells you what to say. None of them check if it's true.

Shark Room is the first pitch simulator that fact-checks your claims against live web sources before the judges score you — so you find out your $30B market size is fabricated before an investor does.

Upload your pitch video. Shark Room verifies your numbers with Google Search, then routes the results to five AI judges (each a named persona with a distinct scoring lens) who evaluate your pitch knowing exactly which of your facts held up. Then it puts you through a live Q&A with webcam recording and real-time eye contact tracking.

Screenshots

Step 1: Add evaluation context (optional — paste competition rules or upload PDF)	The verdict is in — overall score, one-line verdict, and all 5 judge panels at a glance
Named persona scoring — Mark Cuban breaks down each criterion with his characteristic voice	Fact check — claims verified against live Google Search sources, disputed claims flagged with citations
Live Q&A — judge asks a pitch-specific question, you record a 30-second video answer	Session history — on/off topic verdict, judge feedback, and options to retry or continue

---

Practicing a pitch is hard. Friends are too nice. Advisors are too expensive. ChatGPT just agrees with you. And nobody checks whether the market size you cited is actually real.

What Makes This Different

Most AI pitch tools generate generic feedback. Shark Room does three things no other tool does:

1. Fact-checks your claims in real time — every verifiable number is checked against live Google Search before judges see it. Judges know which facts survived.
2. Grounds End User feedback in real community research — Gemini searches Reddit, forums, and review sites to find what your actual target users say about products like yours. The End User judge reacts to real sentiment, not AI guesses.
3. Named personas with distinct scoring lenses — not "an investor" but Mark Cuban, Barbara Corcoran, or Kevin O'Leary, each with their documented criteria and characteristic voice.

What Shark Room Does

1. Upload your pitch video — Shark Room sends it to Gemini for full analysis: transcript, content, delivery, body language, and who your end users actually are.

2. Fact-check your claims — Every verifiable number and market claim is checked against live Google Search results. Judges know which of your facts held up and which didn't before they score you.

3. Face the judges — Five AI personas evaluate your pitch from their own angle:

   • Investor (Mark Cuban, Barbara Corcoran, Kevin O'Leary) — market, moat, ask
   • CFO (Kevin O'Leary, Warren Buffett) — financials, unit economics, return
   • CTO (Jensen Huang, Elon Musk, Linus Torvalds) — technical feasibility, scalability
   • End User — AI-inferred from your video, grounded in real community research
   • Presentation Coach (Vinh Giang, Simon Sinek) — delivery, vocal presence, structure

4. Q&A practice — Each judge asks you a question drawn from their unresolved concerns. You record a 30-second video answer. The coach tracks your eye contact in real time using MediaPipe. Judges evaluate whether you actually answered the question.

5. Bring your own criteria — Upload competition rules or judging criteria as context. Judges will score against those standards instead of generic ones.

---

Stack

• Framework: Next.js 16 (App Router, Turbopack)
• AI: Google Gemini 3.1 Flash Lite — video analysis, judge personas, fact-check, Q&A evaluation
• Grounding: Google Search (via Gemini grounding API) — fact-check + customer voice research
• Gaze tracking: MediaPipe FaceLandmarker — real-time eye contact during Q&A recording
• Styling: Tailwind CSS + shadcn/ui, dark studio theme
• Validation: Zod schemas on all LLM outputs

---

How to Run

# Install
npm install

# Add your API key
cp .env.example .env.local
# Fill in GEMINI_API_KEY

# Dev
npm run dev

# Production
npm run build && npm start

Or double-click start.bat (dev) / start-prod.bat (production).

---

Demo Mode

Set DEMO_MODE=true in .env.local to bypass Gemini and return a pre-built golden fixture instantly. Use this for demos when you can't risk API latency.

---

Built with Cursor

We used Cursor as a full engineering partner, not just autocomplete:

• .cursor/rules/ — project, stack, style, and agent-usage rules loaded into every session so the agent stayed on-brand across 100+ edits.
• Parallel agents — backend pipeline (gemini.ts, API routes, prompts) and frontend (verdict-results, qa-panel) developed in parallel Cursor windows on separate branches, merged at integration points.
• Plan-first workflow — every non-trivial feature started as a plan reviewed before execution. No vibe-coding into unknown territory.
• Multi-file refactors — Cursor handled the full persona system refactor (5 roles × 9 personas × criteria scoring) as a single coordinated edit across 15+ files.
• Iterative prompt engineering — judge persona prompts were written, tested, and refined through Cursor agent sessions with live Gemini output as feedback.

---

Team

• Felix Gabriel Montanez — Frontend, UX, studio theme, component architecture, demo mode
• Xujun Wang — Backend, AI pipeline, Gemini integration, MediaPipe, token tracking

---

Built at Cursor Hackathon SAIT, May 2026.