Common Ground

Inspiration

A group can agree on a place with a high average score while one person's preferences are poorly represented. Common Ground makes that compromise visible.

What it does

Two to four participants choose one to three of six demonstration cultural seeds. The workflow discovers Houston places through Qloo, ranks the shared shortlist separately for each participant, then combines the lowest relative rank with the mean. Users inspect evidence, change the fairness weight, veto a place or revise their interests. Ambiguous names require explicit choices; selected entities are revalidated against current Qloo results.

How we built it

A browser frontend calls a Node.js Docker backend on Render Free. The backend runs fixed, bounded Qloo recommend/rank workflows using a licensed pinned exec-only slice of Qloo's harness. The event API key stays server-side. Workflows and upstream requests are serialized; shortlist identities are batch-validated; repeated identical workflows are cached only within a plan. Provider failures never silently become fixture results. This is a deterministic agentic workflow, not an LLM conversation or a measured claim of outperforming an LLM.

Challenges and lessons

Exact names can identify different entity categories or distinct works. We scoped explicitly typed music and film inputs, retained genuine ambiguity, and added user clarification with stale-choice rejection. Live tests also revealed request amplification and an upstream rate limit. We reduced redundant requests and added bounded, sanitized error diagnostics without automatic retry. Offline fixtures could not establish live API compatibility; we tested the real integration separately.

Verified results

Tested public commit: d634e46632767eed98ed03d6965b2b5424de614f. On October 30, 2026 the default hosted browser flow returned ten real Qloo candidates. Explicit Dune 2021 movie selection returned eight. Hosted stale-choice rejection, veto/cutoff revision, cancellation and subsequent recovery passed. Nine anonymous access/security checks passed. Offline evidence comprises 63 fresh passes, a separately retained unchanged elevated Windows symlink assertion, and 15 browser scenarios plus the final wording check.

Limits and next steps

Fixture rehearsal uses fictional venues and authored scores and is labelled separately. Rank points are relative shortlist positions, not probabilities of enjoyment. Prices, opening hours, bookings, accessibility and availability are not verified; dollar-budget filtering is disabled in live mode. For judging, the owner approved a one-millisecond counting window that effectively removes the app hourly ceiling while retaining single-plan concurrency and provider controls. Render Free cold starts and provider quotas still apply; anonymous usage can consume allocation. Restoration of the original hourly window is scheduled for November 6 at 04:45 UTC and must be verified. Future work includes broader cultural inputs, validated practical venue constraints and a provider-approved judging access arrangement.

Built With

• docker
• javascript
• node.js
• playwright
• qloo
• render