The map on your phone is lying to you, just a little. A restaurant that closed last month still shows five stars and an "Open until 10." A dentist who moved across town is pinned to an empty storefront. Nobody did this on purpose. The physical world simply changes faster than anyone can write it down - and for the companies that sell location data, that decay is an existential problem.
Reprompt, a company from Y Combinator's Winter 2024 batch, exists to fix the part everyone finds boring: checking whether a place is still what the database says it is. Its product is a fleet of AI agents that read the open web the way a diligent researcher would, then update location records for the mapping companies that power everyone else's apps. Foursquare, TomTom, and Radar are among the customers.
01 / The ProblemThe last mile of a map is a person, not a satellite
Satellites and street cars can trace roads and rooftops. What they cannot see is whether the taqueria on the corner still takes cash, changed owners, or shut down over the winter. That "last mile" of location data - the living facts about a business - has traditionally been maintained by hand. Foursquare built a global community of human editors called Placemakers to do exactly this, one verified edit at a time.
The trouble is arithmetic. With hundreds of millions of businesses and dozens of changing attributes each, no human team can re-check the world on any reasonable schedule. The moment data is collected, it starts going stale.
Mapping companies like Radar and TomTom use our agents to get live information about a point of interest or business that keeps their maps up to date. Reprompt, on what it does
02 / The ApproachTeach the AI to distrust itself
Plenty of tools can scrape a web page and guess at a fact. Reprompt's design choice is the opposite instinct: its agents are built to withhold judgment. Before accepting any single attribute - a phone number, an operating status, an owner - the system looks for corroboration from three separate sources. One blog post is a rumor. Three independent references is a fact.
That skepticism runs on an internal agents platform with a few specialized workers. Web agents comb an index of embedded sites with a Perplexity-style search. Vision agents read images and storefronts. Document-reading agents parse filings and registries. A self-driving browser agent is in development. Together they turn research that used to eat roughly 20 workdays into a run that finishes in under an hour.
03 / The FoundersA mapmaker and a grower walk into a startup
The founding pair reads like a deliberate hedge. Lukas Martinelli, the CEO, spent seven years at Mapbox as General Manager of its Search business, running engineering, product, and operations, and oversaw a 120-person team that labeled and evaluated map data by hand. He knows precisely what accurate location data is worth, and how painfully it is made. Before Mapbox he wrote open-source mapping tools that collected more than 15,000 GitHub stars and ended up inside products at Amazon, IBM, and Bosch.
His co-founder and CTO, Rob Balian, comes from the growth side. He led a 50-person Growth team at Robinhood through Covid, the GameStop frenzy, and the IPO, was an early product manager at Facebook, and has trained and shipped vision models. One founder knows the terrain; the other knows how to make a system scale fast.
Before Reprompt, Martinelli's open-source mapping tools earned 15,000+ stars and shipped inside Amazon, IBM, and Bosch. He left the polished part of the map to go build the tedious part. On the road to Reprompt
04 / The CustomersWho buys "is it still open?"
The first buyers are the mapping companies themselves. In 2025 Reprompt became what Foursquare called the first LLM-based Placemaker agent on its Places Engine - an AI editor working alongside the human community rather than replacing it, verifying changes while keeping human oversight in the loop. TomTom and Radar use the agents to keep their points of interest current. Data providers such as Data Axle and Matador appear among its logos.
From there the same engine points at adjacent problems. Fraud and KYB teams need to confirm a business is real and registered; Reprompt runs automated tax and corporate-registry checks globally. Delivery operations need to know where the front door actually is; Reprompt is building entrance-level data for last-mile and autonomous delivery. Site-intelligence customers get brand-level coverage across 100,000+ verified chains.
05 / The MarketInfrastructure hiding under everyone's map
Location data is a crowded shelf - Foursquare's own operations, the open Overture Maps datasets, legacy aggregators, and a wave of AI startups. Reprompt's wedge is not a bigger one-time dataset. It is the promise of freshness: agents that continuously re-verify rather than collect once and hope. That reframing - treating a map as something that decays, and selling the maintenance - is the durable part of the pitch.
The business model follows the data. Reprompt licenses verified location data and enrichment through an API, and sells custom deep-research runs, with revenue scaling by volume and attribute complexity. Backed by Y Combinator and seed investors including Lionheart Ventures, Rebel Fund, Ritual Capital, and Singularity Capital, it remains a small, senior team - the kind that ships infrastructure quietly and lets the customers do the talking.
The most valuable data isn't the data you collect once. It's the data you keep true. The Reprompt thesis, in one line
None of this is glamorous. There is no consumer app, no viral demo. There is a slightly wrong map, a fleet of skeptical agents, and a rule that a fact has to earn its place three times over. For the companies whose products break when a pin points at an empty lot, that is the whole game.