Inspiration
Group plans often reflect the loudest voice. TasteBridge asks a different question: what can everyone reasonably enjoy, while respecting each person's budget, access needs and vetoes?
What it does
TasteBridge is a live Qloo-powered group decision tool. A host reviews one cultural preference for each of 2–6 guests, reviews 2–5 real places, and supplies current price and accessibility facts. Hard constraints and vetoes are applied before Qloo affinity ranking. The tool compares within-query midranks, prioritizes the lowest guest rank and then the mean, and makes ties explicit. People make the final decision; the application does not book or pay.
Try the live app: https://tastebridge-carlosjv.zrm7tv2zmj.chatgpt.site
For a walkthrough without consuming API calls, open the clearly labelled recorded case. Its dated Qloo evidence is real; its test preferences and prices are synthetic operator inputs. It is not a fresh recommendation.
Try it and inspect the evidence
For a no-call walkthrough, open the recorded example, compare its tied choices, then apply a veto to see the recovery guidance. This is dated evidence, clearly labelled.
For a live desktop walkthrough, search and confirm Arrival (2016) and Spirited Away (2001). Add The Museum of Modern Art at 11 W 53rd St and The Metropolitan Museum of Art at 1000 5th Ave, New York. Review current admission facts from the venues and enter the prices, sources and a common USD budget before comparing. In our 28 October test, both general adult prices were USD 30 and the test budget was USD 50; accessibility remained unknown and was not required. Inspect the returned Qloo evidence. Applying a veto must clear the proposal; re-reviewing the changed place and comparing must request another eligible option, without another affinity call. Preferences and budget are operator test inputs, not a real group study.
How we built it
The public implementation uses React/Vinext, TypeScript, Cloudflare Workers and D1. Server-side Qloo Search resolves cultural titles and places; Qloo Insights supplies cross-domain affinity evidence. Address and description support explicit entity review. Signed selections expire after one hour, host-reported facts after 24 hours, and edits invalidate previous proposals.
TasteBridge is a bounded deterministic agentic decision tool with search/evaluate actions. It does not include a separately deployed LLM agent. Scores from separate queries are not assumed comparable: the declared ranking policy uses ordinal ranks within each guest's common shortlist. Missing, incomplete or constant evidence triggers abstention.
Challenges we addressed
Live testing exposed identically named museums in different countries, incomplete affinity coverage and a hosted transport incompatibility. We expose place details, preserve abstention and reject redirects. We also bound request/response sizes, use timeouts, keep the Qloo key server-side, and prevent stale asynchronous responses from appearing after input edits. When constraints leave too few eligible places, the current version explains which facts or options need attention without silently relaxing a budget, access need or veto. Stable guest identities keep each person's entered preference attached to them when another guest is removed.
Verification and trade-offs
A hosted HTTP workflow completed four real Qloo searches and two real Insights requests. The two test profiles disagreed, producing an explicitly tied shortlist rather than false consensus. Anonymous homepage/API access was verified. The earlier release passed ten hosted policy checks and 52 Python regression tests.
An exhaustive synthetic policy comparison covers 216 strict ordinal profiles with three guests and three options. Max-min improves the minimum rank in 27 cases versus mean-only, with an average mean-rank sacrifice of 0.013889. This illustrates a policy trade-off; it does not demonstrate higher actual satisfaction or superiority over LLM recommendations. Desktop browser checks also verified the recorded example, correct tie handling, proposal invalidation after a veto, and abstention when only one option remains. A recorded-case layout check at 390-pixel width was also completed; full mobile usability remains unverified.
The deployed v13 fixes a reproducible guest-removal defect: after entering Arrival / Spirited Away / The Matrix and removing the first guest, the remaining inputs retain Spirited Away / The Matrix; a newly added guest starts empty. The identity regression improved from 2/6 passing expectations on v12 to 6/6 on v13, and the corrected behavior was checked in the production browser. Six offline policy, ranking, recovery, search and mocked-route test commands passed. Independent reconstruction of the published source passed 19 integrity and negative checks and reproduced all 144 source files and the deployed source ZIP checksum. These are bounded engineering checks, not a measurement of recommendation quality.
On 28 October 2026, a fresh end-to-end desktop browser test completed four real searches and two uncached Insights requests. Both test profiles ranked The Metropolitan Museum of Art above MoMA; the displayed proposal matched those per-guest affinities. A veto immediately removed the proposal, and the blocked comparison requested another eligible place without increasing provider usage. The durable application counter increased from 8 to 14 of 60 during the live test, leaving 46 calls at 00:48 Europe/Madrid. This verifies that run, not future availability, real-user outcomes or a guaranteed recommendation. No reservation or payment was made.
On 30 October 2026, deployed v14 adds a read-only readiness panel before the live form. It distinguishes configuration presence and known, unknown or exhausted shared allowance without contacting Qloo. It links directly to the dated recorded example. The production panel showed 46 of 60 calls remaining at the time checked; this is a snapshot, not a reservation or a provider-health test. The change passed 22 readiness checks, eight component checks, a production build and real-browser hydration/refresh checks. A 390-pixel frame layout check showed no horizontal overflow; full mobile usability remains unverified. The published 147-file source snapshot reconstructs the deployed source archive exactly, with 17 additional packaging and rejection checks. These are software checks, not measured recommendation quality.
What we learned
Cultural intelligence is most useful when a decision process can explain uncertainty and disagreement. Qloo affinity does not establish venue price, accessibility or personal enjoyment. Those boundaries must remain visible.
Availability and next steps
The experimental public demo enforces a durable total allowance of 60 provider calls with a one-hour cache and no automatic retries. This is our application cap, not a statement of Qloo's provider quota. Live requests can exhaust the allowance; the recorded evidence walkthrough remains available without live calls. Provider credential expiry and judging-period allowance still require confirmation. The rules require unrestricted free testing through judging; the shared cap and recorded fallback do not establish that this condition is satisfied.
Next steps are usability testing with real groups, broader venue coverage and comparative recommendation evaluation. No customers, revenue or measured satisfaction gains are claimed.
Open source and development disclosure
Developed by Carlosjv with AI assistance. Complete current hosted v14 source, assets, tests, dependency locks and retained vendor notices are published under MIT at https://github.com/carlosjunquerovila-afk/tastebridge. From a clone or download, run python3 reproduce_hosted_v14.py and follow the README for tests and build. The self-contained full source patch and checksum inventory reconstruct the deployed version; the historical 0.3.2 archive is not required. No credentials or runtime data are included. Earlier synthetic and Python prototypes remain clearly separated and retained as history.
