Inspiration
Choosing a film together is a negotiation. A useful recommendation workflow should show where people overlap, preserve disagreement and leave the final decision to the group. Taste Bridge explores how to make that evidence visible without pretending that a ranked list knows what every person wants.
What it does
The published application has two clearly separated experiences. A recorded-result view shows ten real public movie recommendations previously retrieved from Qloo for Inglourious Basterds. A separate three-person comparison illustrates explicit film confirmation, candidate coverage, individual ranks, partial overlap and failure recovery using invented inputs and synthetic or labeled replay data.
The real-result view is labeled “Replay of previously retrieved real API results.” It identifies the input and retrieval time. Clicking “Show recorded results” displays ten movies including Pulp Fiction and Django Unchained. Viewing makes no new Qloo request and does not generate live or individually personalized recommendations.
How we built it
The application uses JavaScript, Cloudflare Workers, Static Assets, and a SQLite-backed Durable Object. Custom HTTP function tools, resolve_movie and compare_confirmed_movies, validate explicit choices and expose a callable retrieval-and-comparison workflow. Ranking is deterministic. We do not claim an external LLM, agent framework or MCP server. AI coding assistance was used.
An official Qloo Insights request using a fixed public movie identifier on October 31 returned HTTP 200, valid JSON, success=true, and ten validated records in a 443,671-byte response. The request used a finite 2 MiB streaming limit, an eight-second deadline, and no automatic retries or redirects. Credentials remain server-side and are absent from the browser and repository.
Challenges
A 256 KiB cap initially stopped the payload before JSON decoding. Diagnostics identified the issue, and a finite-size repair enabled the verified response. Persistent reservations preserve consumed attempts, including uncertain failures; further sending was stopped. A browser initialization race was also fixed.
Validation passed 59 Worker tests, 21 public HTTP checks, and 16 desktop/mobile Chromium checks.
What we learned and current limits
HTTP 200 alone is not enough: response shape, records and provenance need separate verification. This demo is a fixed recorded observation. Visitor-entered films do not trigger Qloo requests. The invented comparison does not establish three independent live preference profiles. There are callable tools, but no demonstrated autonomous agent or external model client.
We do not claim that this replay scope necessarily satisfies the hackathon's end-to-end, Qloo-powered or agentic requirements. A next iteration would independently verify per-person result lists under bounded access.
Code is MIT-licensed; Qloo data and marks remain subject to applicable terms. We make no unmeasured quality, CPU or free-quota claims.
Try it
Open https://tastebridge-demo.yagi-5e3.workers.dev/live without a login or API key. Read the replay disclosure, then click “Show recorded results” to see ten movies starting with Pulp Fiction. Replaying does not create duplicates or make a new Qloo request.
The separate invented comparison uses Quiet Harbor, Orbit and Forest. Resolve the inputs, inspect Orbit's two invented choices, explicitly confirm three choices, and compare. This does not personalize the ten real results. The root demo provides synthetic-fixture checks for partial overlap, failure and recovery.
Source: https://github.com/aisoh877/tastebridge-qloo For local testing use Node 24.18+ and Wrangler, then npm run check, npm test and npm run dev. Live/operator/grant access is disabled in the source.
Log in or sign up for Devpost to join the conversation.