Zemi-Aku

Try it in 60 seconds

Open https://zemi.r21digital.com/app → choose Bakeries → tap Río Grande. Read the verdict, its sources, what the data cannot tell you and the field checks. The Opening plan to print link turns it into one page to carry on a visit: https://zemi.r21digital.com/plan?industry=bakery&town=R%C3%ADo%20Grande&lang=es

Explore an artist's audience affinities, inspect the Evidence drawer and visit https://zemi.r21digital.com/compare for saved with/without-Qloo comparisons. Choose Español on the first screen for the session in Spanish.

Current availability: The opportunity map, town verdicts, printable opening plans and rule-based answers work without a model turn. Live model answers are available under daily and per-visitor limits; the Evidence drawer identifies live execution, tool calls and timing. Use “Ask again without Qloo” to compare the same question with Qloo tools removed. The Evidence line identifies the answer engine. The comparison page contains saved runs.

Inspiration

Before signing a lease, which town deserves a closer look? Puerto Rico's 78 municipalities have different cultural tastes and different business landscapes. A list of popular places does not explain that difference. Zemi-Aku puts Qloo's taste signals beside public business and population data so an entrepreneur can compare towns and inspect the evidence.

What it does

Choose cafés, bakeries, restaurants, bars, hotels or salons. The opportunity map compares Qloo taste with Census establishments per 10,000 residents across the island. Open a town, read the sources, then change the industry and compare again.

The map keeps two axes visible. It does not turn them into a prediction of demand or profit. Census counts below three are unpublished; those cells appear as 0–2, not a confirmed zero. Qloo catalog counts, Overture observed supply and Census estimates retain their own labels, dates and limits.

Every town verdict links to a printable opening plan: the verdict, a comparison town, places to visit with dated hours, a field checklist, limitations, sources and a QR code back to the live verdict.

Music offers another way to explore: an artist's audience affinities lead to real Qloo places. An affinity is not a visit or endorsement. Open a place record and inspect its surroundings on the 3D map. Speech playback has ten voices across five languages.

How it was built

A browser JavaScript client and Cesium map connect to one Node server. Qloo supplies catalogs, artist/place affinities, tag search, audience information and heatmaps through REST endpoints including /v2/insights and /search. The server also uses Qloo's official local MCP harness. Open data adds separately labeled Overture places, Wikidata landmarks, Census context and weather.

The open-question path selects from 18 read-only tools. Only validated tool results control map actions. The Evidence drawer exposes tool calls, arguments, sources, counts, timing and resolved Qloo IDs. Deterministic research answers and saved replays carry different labels. Provider or budget failures use the documented fallback paths.

Daily budgets, visitor limits and session reservations bound the public demo. The hosted MCP connector remains off pending written clarification from Qloo; the local path uses the developer's own key.

The published /compare page keeps seven saved pairs. The seventh keeps Qloo's place catalog and removes only the taste tools, isolating the affinity signal. These show Qloo's contribution on those questions; they are not a general accuracy benchmark.

Challenges and lessons

A source can know a place without proving it belongs to the municipality a visitor named. Boundary checks and source labels matter. So do sparse coverage and Census suppression. The interface carries those limits alongside the result.

Speech and asynchronous map actions add another challenge: an old response must not replace a newer question. Cancellation, response ownership and relay limits address those paths.

Evidence and limits

A preregistered backtest on Census 2019→2023 counts found that gap towns gained more businesses than equally sparse, lower-taste towns in five of six industries. The result was not statistically significant (p = 0.07); bakeries were the one loss. This does not establish demand, profitability or a causal effect.

Backtest: https://github.com/CDVolvik/zemi-aku-public/blob/main/docs/BACKTEST-GAP-2026-10-26.md

The latest release has 525 automated tests. The public snapshot passes 521 and skips four that require private data. Independent owner feedback and real-phone usability checks remain pending; no user quote or business outcome is claimed.

What's next

Complete real-phone checks and obtain feedback from Puerto Rico business owners with written permission before quoting them. No native iPhone app has been released.

Live project: https://zemi.r21digital.com Code: https://github.com/CDVolvik/zemi-aku-public