Lost Dutchman's Compass

Inspiration

A favorite film, book or musician can be a starting point for discovering a place. Lost Dutchman’s Compass explores that idea for first-time visitors to the Mesa area, with a focus on local history, documented legends and nature. It connects cultural taste to a small, transparent choice set instead of inventing an itinerary or presenting local lore as established fact.

What it does

Visitors compare two to four places from a six-place catalogue across Mesa, Apache Junction and Superior. They resolve and explicitly confirm each place and one or two cultural favorites, then ask Qloo to rank the confirmed choices. Returned cards retain Qloo’s order and affinities. Omitted choices remain visibly unranked, and missing explanations are not filled in with invented reasons.

Optional History, Legends and local lore, and Nature and outdoors interests provide independently sourced catalogue guidance. They do not alter Qloo’s scores or send extra provider requests. A local shortlist helps the visitor decide what to investigate next; official venue links are the source for current visitor and access information. This is a comparison aid, not route planning, a booking service or a guarantee of enjoyment.

How it was built

The desktop web app uses JavaScript, HTML and CSS, a Node.js server and the official @qloo/qloo-harness CLI, pinned to version 0.1.26. It is packaged with Docker and hosted on Render. Qloo powers cultural-entity resolution and taste-based place ranking. The server keeps provider authentication out of the browser and bounds each explicit operation without automatic provider retries.

The agentic route is an external browser-capable agent operating the visible controls: checking identities, choosing the intended matches, requesting a bounded ranking, inspecting partial coverage and selecting a provisional shortlist. The public app also supports a human using those same controls. It does not contain an in-app conversational model or autonomous itinerary generator. The source includes a reproducible tool-use workflow and distinguishes agent decisions from Qloo output.

Challenges and lessons

Identity matching needed explicit confirmation of names, addresses, types and the intended cultural work or version. Session-scoped identities expire after 15 minutes. Independent review found same-title favorite deduplication and redaction-boundary issues; regression tests now cover those cases. Busy states, bounded dispatch accounting and one live submission at a time protect the free hosted instance.

The important lesson is to preserve uncertainty. A documented hosted rehearsal returned two of three submitted places; the third stayed unranked. An earlier blinded comparison was all ties. Those results demonstrate a working partial workflow, not recommendation-quality superiority or complete coverage. The existing approximately 1:36 actual-app walkthrough, with narration and captions, was independently reviewed; it is separate from the earlier dated rehearsal metrics.

Accomplishments

• A working public desktop demo with real Qloo integration through the official CLI.
• Explicit resolve → confirm → rank controls, local shortlist interactions and per-operation evidence.
• Independent source, adversarial and walkthrough reviews. The original full offline suite passed 149 tests; the later source-publication check passed all 140 dependency-free tests, plus 20 focused checks across four rewritten historical snapshots.
• MIT-licensed public source. Historical response-derived fixtures and mappings were cleaned from reachable main history while preserving original commit authors, dates and structure, with a privately retained, restored-and-verified rollback. This does not claim erasure of old caches or previously copied objects.

Try it and judge the result

1. Open the public demo and choose “Build a broader Mesa-area comparison.” Wait for readiness; the free service may need time to wake.
2. Choose two to four places. Resolve each one, inspect its identity and explicitly confirm the intended match. Choose “None of these match” when appropriate.
3. Resolve and confirm one or two cultural favorites, then select the confirmed favorites to use.
4. Click “Rank confirmed choices” once. Read coverage, returned order, missing explanations and “Inspect this operation’s evidence.” Ranking allows at most seven provider dispatches.
5. Build a local shortlist and use the official venue links to check current conditions. “Clear local selections” resets the comparison without a Qloo call.

Wait through displayed busy or rate-limit pauses before a new explicit attempt. Re-resolve expired identities. Do not silently substitute unmatched entities or rerun an evaluation to obtain a better outcome.

Live demo: https://lost-dutchmans-compass.onrender.com

Source and complete judge instructions: https://github.com/scking21/Lost-Dutchmans-Compass

What’s next

Broader real-visitor evaluation, better evidence about coverage and usefulness, and clearer recovery from missing matches. Any future itinerary or conversational layer should preserve the same confirmation, provenance and uncertainty boundaries.

Built With

• css3
• docker
• html5
• javascript
• node.js
• qloo
• render