PitchCraft Agent

🛡️ TicketGuard

Spot ticket-resale scams before you pay.

An AI agent that investigates a suspicious ticket listing or seller DM and returns an evidence-backed scam-risk verdict — grounded in MongoDB, reasoned by Gemini.

[](https://frontend-nu-ochre-z41mw3z0l5.vercel.app)

---

The problem

Major-event ticket resale is a fraud magnet — the FBI, FTC, and BBB are actively warning about a 2026 World Cup resale-scam wave. Buyers have no way to tell a real listing from a "pay me on Zelle, I'll email the PDF" trap until the money's gone.

TicketGuard is a consumer agent that runs a real multi-step investigation on a pasted listing / DM / screenshot / URL and returns a verdict: SCAM · SUSPICIOUS · LIKELY-LEGIT — with a confidence score and cited evidence. It never claims a ticket is "authentic" (no third party can); it speaks only in risk signals.

> Built for the Google Cloud Rapid Agent Hackathon 2026 — MongoDB track.

---

🏗️ Architecture — MongoDB is the brain

flowchart LR
  U["👤 Paste DM · upload PDF/image · drop URL"] --> FE["Next.js UI(Vercel) · live SSE stream"]
  FE -->|"POST /api/investigate"| BE["FastAPI + Google ADKGemini 2.5 Flash"]

  subgraph PIPE["8-step agent investigation"]
    direction TB
    N["1· Normalizer"] --> R["2· Hybrid Retrieval"] --> REP["3· Reputation"] --> F["4· Forgery / Duplicate"] --> S["5· Risk Scorer"] --> RULE["6· Transfer Rules"] --> V["7· Verdict"] --> P["8· Persist"]
  end
  BE --> PIPE

  R |"$vectorSearch + $search → $rankFusionvia MongoDB MCP server"| DB[("MongoDB Atlas")]
  REP  DB
  F |"sha256 barcode dedup"| DB
  S |"$group / $facet scoring"| DB
  P --> DB
  DB -.->|"change stream"| FE

  classDef mongo fill:#00684A,color:#fff,stroke:#00684A,stroke-width:2px;
  class DB,R,REP,F,S,P mongo

Every green node is real MongoDB work. The risk score is computed in the database ($group/$facet), not by the LLM — so it's reproducible and auditable. The verdict writer only explains the gathered evidence.

---

🏆 How it meets the hackathon bar

| Requirement | TicketGuard | |---|---| | Gemini agent, multi-step mission | 8-step ADK pipeline (not a chatbot): normalize → retrieve → reputation → forgery → score → rules → verdict → persist | | Integrates the MongoDB MCP server | Agent reaches Atlas through the official mongodb-mcp-server (read-only find/aggregate/count/collection-schema) | | Deep, meaningful MongoDB use | Vector Search + Atlas Search hybrid ($rankFusion on 8.1+, code-fusion fallback) · server-side $group/$facet scoring · change streams for the live feed · barcode-hash dedup · agent memory | | Human-in-the-loop | Report / Find verified resale / Proceed — and Report writes to Atlas → change stream → the global corpus gets smarter | | Multi-modal ingestion | paste text · upload PDF/screenshot (Gemini vision OCR + tamper hints) · paste a URL | | Runs on the web, hosted, open-source | Next.js (Vercel) + FastAPI (Cloud Run / Render), MIT licensed | | Honest by design | No mock ever drives a verdict — DB-dependent steps return not_configured rather than faking |

---

🔎 The 8-step investigation

| # | Step | Engine | MongoDB | |---|------|--------|---------| | 1 | Normalizer | Gemini structured extraction | — | | 2 | Hybrid Retrieval | $vectorSearch + $search, fused | scam_corpus | | 3 | Reputation | typosquat distance + prior reports | reports | | 4 | Forgery / Duplicate | sha256(barcode) lookup + PDF/image tamper hints | tickets_seen | | 5 | Risk Scorer | server-side $group/$facet (not the LLM) | Atlas | | 6 | Transfer Rules | deterministic rule engine | official_rules | | 7 | Verdict | Gemini, grounded only on evidence | — | | 8 | Persist | agent memory | investigations |

Real verdict from the live API (Atlas off → DB steps honestly not_configured):

{ "verdict": "SCAM", "confidence": 0.9,
  "evidence": ["Unofficial PDF transfer", "Irreversible Zelle payment",
               "Price 120 USD far below face 350 USD", "Urgency cues", "Violates official transfer rules"],
  "engine": "gemini-2.5-flash · Google AI Studio" }

---

🧩 Tech stack

• Frontend — Next.js 14 (App Router) + Tailwind + Framer Motion, light/dark, consumes SSE. Deploys to Vercel.
• Backend — Python FastAPI + Google ADK agents, Gemini 2.5 Flash (AI Studio or Vertex), streams SSE. Deploys to Cloud Run / Render (Docker ships Node + Python so the MCP server runs).
• Data / brain — MongoDB Atlas: Vector Search (gemini-embedding / text-embedding-004, 768-dim) + Atlas Search, change streams, the official MongoDB MCP server.

---

🚀 Quick start

# Backend
cd backend
python -m venv .venv && .venv\Scripts\activate      # (Windows)
pip install -r requirements.txt
copy .env.example .env                               # fill MONGODB_URI + GOOGLE_API_KEY
python scripts/setup_atlas.py                        # create indexes + seed corpus
uvicorn main:app --reload --port 8001

# Frontend (new shell)
cd frontend && npm install && npm run dev            # http://localhost:3000

Full setup, deployment, env vars, and how to edit the Gemini prompts → [HANDOFF.md](HANDOFF.md).

---

📊 Evaluation

A labeled set of synthetic scam/legit examples (backend/data/eval_set.json) measures precision / recall so impact is quantified, not claimed.

⚖️ Responsible-use

Decision-support only — not a guarantee; verify independently. Risk-signal language, never accusations. Synthetic demo data only. No trademarked marks.

📂 Structure

backend/   FastAPI + ADK agents + MongoDB (agent.py · pipeline.py · ingest.py · db.py · scripts/setup_atlas.py)
frontend/  Next.js UI (app/ · components/ · lib/)
HANDOFF.md Full contributor + deploy guide

License

[MIT](LICENSE)

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vaibhav4046
• Source: vaibhav4046/PitchCraft-Agent
• License: MIT
• Homepage: https://frontend-nu-ochre-z41mw3z0l5.vercel.app

Install and usage instructions live in the source repository linked above.