Screenscore

• 

  Landing page

• 

  Pipeline

• 

  google adk

Inspiration

We kept seeing AI agents confidently make up movie ratings and box office numbers. For real acquisition decisions, that's useless. We wanted something that can't lie — every figure has to come from a live database query.

What it does

You give ScreenScore a movie title. It queries 388K IMDb titles in ClickHouse, runs real SQL, and gives you an ACQUIRE / PASS / FURTHER REVIEW recommendation with genre benchmarks, director track records, and risk flags.

How we built it

• Google ADK + Gemini 3.1 Flash Lite
• ClickHouse via MCP (388K movies, 817K actors, 3.4M roles)
• 8-step deterministic pipeline
• Anti-hallucination self-audit that fails the pipeline if data doesn't match
• Cloud Run with auto-deploy from GitHub

Challenges we ran into

• The LLM kept hallucinating. We'd give it SQL results and it'd still report wrong numbers. Built a strict self-audit gate to catch it.
• Free tier quota is tiny. Had to cut our prompt by 51% to fit.
• MCP fails silently if env vars are wrong. No error, just no tools.

Accomplishments that we're proud of

• The anti-hallucination system actually works — tested it hard, catches fakes every time
• 115 tests passing
• The memo output is legit — not just a chat summary

What we learned

• Grounding in real databases changes everything. The agent doesn't "know" movies — it queries them.
• Prompt engineering is half the project. Seven iterations before it worked within quota.
• MCP made database integration take hours instead of days.

What's next for Screenscore

• Persistent sessions across page loads
• More database sources (box office, streaming data)
• Batch analysis — compare multiple titles at once

Built With

• clickhouse
• docker
• fastapi
• gemini
• googe-adk
• google-cloud
• llm
• mcp
• python