Why Your AI Agent Hallucinates About Real Places

Ask a frontier model for the elevation of a specific New York lot at 40.7128, -74.006 and it will hand back a number between sea level and fifty meters, with no source and no timestamp. Mireye, a Y Combinator (S26) startup, bets the fix is not a smarter prompt but a fresh, county-by-county ground-truth layer that sits under the model.

Mireye ships that layer as a single API and an MCP server that an agent can call. Behind the surface are 366 normalized fields covering raw data, enrichment, tools, and signals, all referenced to specific US addresses. The fields break down into address geometry, parcel ownership, zoning, hazards, building permits, tax, and a job-pricing quote endpoint an agent can hit before scheduling work. An agent pastes a skills.md file and gets back, for example, the 13.03 m elevation at the same New York coordinates, sourced from USGS 3DEP COG, with a fetched timestamp and a confidence score.

Founder Ansh did not start with a thesis. He started with a failure. While building a niche property app, he watched frontier models confidently invent the owner, the acreage, and the zoning of real parcels. A Fortune 500 insurer's engineers told him they had given up on underwriting agents, the automated risk assessment of a property's physical and legal state, for the same reason: every model they tested hallucinated the ground truth about the property. The validation was not theoretical. It was a customer who had already quit.

Of the 317 catalog fields the niche app exposed, 311 got queried in the wild, with no single use case dominating. Ansh killed the app and kept the engine. The insight was structural: agents do not want a feature, they want the whole surface. The 5,000-credit free tier, with no card required, exists to prove the engine gets the same demand from outside his original vertical.

The shipped skills are not generic. A "site screening" skill wraps parcel resolution, drive-time geometry, and zoning lookups. An "address underwriting" skill bundles elevation, parcel data, and a job-pricing quote endpoint. Both are the load-bearing examples of the layer the company is selling: the boring, deterministic plumbing that frontier models will not invent for themselves, even when prompted to be careful.

The hard part is not the API. It is the loop behind it. County assessor data, federal elevation tiles, parcel boundaries, and permitting feeds are gathered and normalized record by record, with provenance, semantics, and refresh cadence contracted per source. That loop has to stay fresh forever, or the layer turns into the same hallucination it was meant to replace. Mireye's bet is that the cost of that loop, run by a small team with deep county-by-county access, is lower than the cost of every underwriting agent rebuilding it from scratch.

The next test is whether the engine keeps its demand surface as it leaves the founder's own use case. Ansh's bet is that any agent asked about a real US address will keep asking the same 311 questions, regardless of who wrote the prompt. The free key and the MCP install path are designed to find out.