PARALLAX

Choose one New York place through the culture you love—with a question worth answering.

Comment

Inspiration

A favorite artist and a memorable film can suggest different ways into a city. Visitors often know their cultural references better than a destination's neighborhoods. PARALLAX explores how those references can become a practical choice while leaving the person in control.

What it does

Confirm two or three artists or films. Qloo discovers New York places and ranks one shared shortlist through each reference. PARALLAX selects a prospective A/B comparison when its answer could help the decision under an explicit finite model. Up to three relative choices refine the result.

The person can accept both, veto neither-for-this-outing options, skip, change an answer or remove it. Retained answers rebuild the model from a clean prior. The studio includes a responsive venue view, geographic map, cultural-connections plot, before/after ordering, evidence, local preferences and field-note export.

The agentic-tool path is parallax_advise: a local MCP stdio bridge to the same public HTTPS backend and decision engine. An agent can resolve references, prepare a candidate set, request the person's choice, advise on explicit answers and forget temporary evidence. The runnable protocol demonstration makes actual tool calls with declared sample input; it does not invoke an LLM.

How it was built

React and TypeScript power the studio. Node functions on Vercel call the Qloo event API. Qloo supplies identity confirmation, place discovery and supplied-shortlist rankings. A balanced common pool has at most ten place identities; missing ranks remain missing.

The decision core enumerates a small simplex of cultural-lens weights and updates it with a logistic choice likelihood. It evaluates the expected reduction in modeled decision regret for eligible comparisons. These are application-defined ordinal utilities, not provider probabilities or measured enjoyment.

An atomic Postgres ledger enforces shared provider allowances; signed sessions and identity proofs protect API boundaries. MapLibre and OpenFreeMap supply geographic context. IndexedDB stores studio preferences. Verified photographs have responsive local WebP editions and individual source/license credit. Runtime operation requires no mandatory paid model or image subscription.

Challenges

The same candidate set must be ranked under every lens for comparisons to be interpretable. Missing results cannot be treated as dislikes. Removing an answer must rebuild the actual model. Serverless quota enforcement must remain shared and fail closed. The experience also needed careful map lifecycle handling, responsive imagery and honest evidence labeling.

What was verified

Eighteen unit/HTTP/asset tests cover an independent finite-model oracle, answer semantics, scoped API boundaries and image assets. Eight browser cases cover the real interface and renderer with synthetic API fixtures. Separate checks exercised actual deployed Qloo retrieval, map controls, correction, restoration, responsive photo delivery and the MCP protocol. The synthetic report includes 1,600 seeded executions and retains unfavorable comparisons. No independent human preference study or universal superiority is claimed.

What is next

Evaluate usefulness with real visitors, investigate broader geographic coverage and improve the model under nonlinear or changing preferences. These are future directions, not completed outcomes. The current edition focuses on choosing one New York place; it does not book venues or certify visiting conditions.

Links and technologies

Built With

  • github-actions
  • indexeddb
  • maplibre-gl-js
  • model-context-protocol-sdk
  • neon-postgres
  • node.js
  • openfreemap
  • qloo-taste-ai-api
  • react
  • typescript
  • vercel
  • vite
  • zod
Share this project:

Updates

Submission history