Launch HN: Mireye (YC S26) – Infrastructure for Physical World AI Agents
Hacker News anshchokshi
Mireye launched an API for AI agents to judge real-world places. It feeds them location data, signals, and tools so they stop guessing about physical sites.
Based on reporting by Hacker News, anshchokshi — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mireye wants to be the plumbing for AI agents that have to deal with the real world. Founder Ansh says the company gives models data, enrichment, tools, and signals for any US location through one API and an MCP server. The pitch is simple: if an agent can talk about the internet, fine. If it has to reason about a parcel, a street corner, or a warehouse, it needs better ground truth.
That came out of pain, not theory. Ansh says he hit the same wall while building construction agents, and a Fortune 500 insurer told him its own underwriting agents had stalled for the same reason. Models keep stumbling when a question is tied to one exact place. So Mireye starts with a bare address and turns it into things like owner, acreage, structures, nearby power, plus alerts when something changes, such as a rezoning filing.
The first product was a site-screening app. It worked well enough that people testing it on places they knew trusted the answers. But they didn’t care about the app. They cared about the engine. Usage backed that up: 311 of 317 catalog fields were queried, and no single use case dominated. So the app got cut loose and the infrastructure became the product.
What makes the system interesting is the refusal to pretend missing data is a clean answer. Mireye says each field comes back as ok, absent, or failed. That matters because the same field name can mean different things in different counties, and a blank can mean “not mapped” instead of “nothing here.” The company also says it does on-demand indexing: if a customer asks for a field it doesn’t have, a long-running agent goes out, researches sources, checks against ground truth, and usually indexes it within a day.
The company says it now handles 366 fields and runs multi-tenant from one index. It also says the hard part is ugly and ongoing: county-by-county collection, normalization, contracts that explain what each source means, and keeping everything fresh. Customers are already using it for flood, wind, and wildfire screening, cleaning messy listing addresses, scoring street corners for poster spots, warehouse sourcing, data center siting, drone planning, school bus routes, and human trafficking investigations. Pricing is public, with a free tier, a $19 plan, a $99 plan, and custom enterprise pricing.
My take — AI-written commentary, not fact-checked reporting
This is the right instinct: if an AI is making decisions about the physical world, “close enough” data is just expensive fiction. The bigger story is that the useful moat isn’t a smarter model, it’s the miserable work of typing absence, chasing county records, and making the machine admit when it doesn’t know. That’s not glamorous, which is exactly why it might actually work.
Read more about this at: Hacker News
