CultureRisk AI

An autonomous agent that uses Qloo to check whether a film's comp audiences really overlap, then files a coverage memo with a 0-100 culture-risk score. Know the audience before you greenlight.

Comment

Inspiration

A studio greenlight commits $30M–$300M, and the core risk model is the comparable-title pitch: "La La Land meets Hereditary". Nobody checks whether those two audiences actually overlap until test screenings, months and millions later. Qloo's taste graph can answer that on day one.

What it does

CultureRisk is an autonomous research agent for film-studio executives. Give it a project and its comps. It plans an investigation across Qloo's insights, demographics, tags, compare and search endpoints, changes course as evidence arrives, and files audience coverage: a script-coverage-style memo that ends in a stamp (Recommend / Consider / Rework / Pass) with a 0–100 culture-risk score and the formula behind it.

  • Live comp picker backed by Qloo search, with real posters.
  • Taste overlap and demographic fit between comps, plus concept fit from the script's own genre tags.
  • Partner explorer for brands, music, locations, TV and podcasts, toggling distinctive vs mass-market, with explainability showing which comp drives each pick.
  • A "popularity trap" section: famous names this audience is lukewarm on.
  • Adaptive branches: if the comps' audiences split, the agent searches for a bridge title and offers a one-click re-run.
  • An "On the desk" panel that streams every step, every Qloo call with its parameters, and every decision.
  • It is also an MCP server, so any agent (Claude, Cursor) can call it as a tool.

How we built it

Plain Node 20 on Netlify Functions, zero runtime dependencies, browser ES modules. Each request runs one agent action and echoes the evidence ledger back to the client, so investigation depth is unlimited within Netlify's 30 s cap. The agent is deliberately deterministic: the same evidence gives the same verdict, no entity can be hallucinated, and the server verifies every recommended ID against what Qloo returned in that session.

Challenges we ran into

Calibrating scores on live Qloo data. Cross-affinity ranges are narrow (Paddington 2 × Saw ≈ 0.65 vs The Revenant × Everest ≈ 0.93), so we calibrated against real pairs. We also had to keep provenance honest: every figure is labelled live, simulated, fallback or mixed, and the memo warns when it isn't fully live.

Accomplishments that we're proud of

  • A research agent with no LLM in the loop: auditable, repeatable, and every claim traces to a logged Qloo call.
  • Provenance you can trust: the memo says exactly how much of its evidence was live.
  • 66 passing tests and zero runtime dependencies.
  • It works both as a human-facing memo and as an MCP tool other agents can call.

What we learned

Qloo's affinity vs popularity split is the most useful signal for finding under-priced partners, and explainability turns a score into an argument an exec can defend.

What's next for CultureRisk AI

Back-test the verdict against known hits and flops, and extend beyond film to music, games and TV.

Built With

Share this project:

Updates

Submission history