PrizePilot AI

Gemini understands the rules. PrizePilot turns missing evidence into an approved, verified fix.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for PrizePilot AI

Inspiration

A working project can still be disqualified when rules, repository evidence, deployment proof, and deadlines are scattered across different systems. PrizePilot exists to close that gap.

Gemini understands the rules. PrizePilot turns missing evidence into an approved, verified fix.

What it does

PrizePilot converts competition rules into structured submission requirements with Gemini, inspects a GitHub repository, connects each requirement to files, URLs, and API observations in an Evidence Graph, and calculates an explainable risk score.

Its lifecycle is:

Discover → Decide → Approve → Act → Revalidate → Prove

Historical verified run — Act → Prove

A verified run detected missing README spin-up instructions, proposed a bounded repair, and—after explicit user approval—created one separate branch, one commit, and Draft PR #1. PrizePilot revalidated the PR branch rather than main: the evidence changed FAIL → PASS and Risk Score 15 → 5. The PR remains open, draft, and unmerged.

Live read-only cloud verification

The frozen Cloud Run deployment is deliberately read-only. In the recorded live run, gemini-3.6-flash returned structured output with fallback=false, extracting 4 requirements. PrizePilot built 4 Evidence Graph nodes, detected missing demo-video evidence, and calculated Risk Score 40 with visible reasons. Firestore stored the real-time analysis and repository-inspection audit events. Cloud mode contains no GitHub token and made 0 GitHub write requests.

These are two clearly labelled evidence lanes: the historical run proves the approved external action; the live cloud run proves current Gemini, ADK, Cloud Run, and Firestore operation without pretending the deployed service still has write credentials.

How we built it

  • Gemini 3.6 Flash on Vertex AI for structured rule extraction
  • Google ADK for agent orchestration
  • FastAPI on private Cloud Run
  • Firestore for audit events and state history
  • GitHub REST API for public repository inspection and the separately verified historical Draft PR workflow
  • A deterministic Evidence Graph and risk engine for traceable decisions
  • Approval IDs and idempotency keys to prevent duplicate actions
  • An offline extractor and local persistence fallback for reproducible testing

Vertex AI uses Application Default Credentials; no API key is embedded in the repository.

Challenges

The hard part was evidence integrity across understanding, action, and revalidation. We had to distinguish base and PR branches, prevent duplicate operations, preserve failed attempts, and avoid claiming more than an audit record proves. That led us to separate historical action evidence from live read-only cloud verification.

Accomplishments

  • A real Gemini call found mandatory technology requirements missed by the offline extractor.
  • A real public Draft PR proves approval-gated remediation without automatic merge.
  • Branch-aware revalidation proves FAIL → PASS and Risk Score 15 → 5.
  • Cloud Run revision prizepilot-ai-00001-lk7 runs the backend with 100% traffic.
  • Firestore records real-time requirements.extracted and repository.inspected events.
  • Cloud mode stores no GitHub token and performs zero GitHub writes.
  • The frozen local regression suite passes 12 tests and Ruff.

What we learned

An LLM answer becomes operationally valuable only when it is tied to inspectable evidence and deterministic controls. Human approval, idempotency, branch-aware revalidation, and honest provenance matter as much as model quality.

Data sources

PrizePilot uses user-provided competition rules and public GitHub repository metadata, files, branches, commits, and pull-request observations. No private user data is required for reproduction.

Reproduce it

The public source repository includes the architecture diagram, license, tests, local spin-up instructions, and a detailed reproduction guide. Local offline analysis requires no cloud credentials. Vertex AI verification uses the evaluator's own Google Cloud project and ADC. The hosted Cloud Run service is private, so the video provides the live cloud proof without publishing credentials.

Disclosure

This project was created during the hackathon submission period. It uses open-source dependencies listed in the repository; development used an AI coding assistant. No undisclosed pre-existing application code was incorporated. The historical approval record is transparently identified as consolidated after the authorized GitHub action; it is not presented as an original live Firestore timestamp.

Built With

  • cloud-run
  • evidence-graph
  • fastapi
  • firestore
  • gemini-3-6-flash
  • github-api
  • google-adk
  • python
  • vertex-ai
Share this project:

Updates