# Finding the Front Door: How Zephr is Collaborating with Overture to Teach AI Agents to Understand the World Around Them > Source: > Published: 2026-07-29 17:39:53+00:00 Walk down Pearl Street in Boulder, Colorado with Sean Gorman and he’ll point out things most people filter out: which storefront signs are actually visible from the sidewalk, which business entrances face the street and which open onto an alley, which coffee shop’s awning is the only thing telling it apart from the office next door. These might seem like tiny details. But for Gorman, a geographer, they’re the raw material of a question that has captivated him for years, and the one his startup, [Zephr](https://zephr.xyz/), exists to answer: how do you teach a machine to understand where it is, and what’s around it? ## Maps were built for humans To see why that’s hard, start with the map on your phone. It’s built for human eyes: a representation of the places, buildings, and streets of the world with a blue dot on it. The blue dot is only approximately you (a phone’s GPS routinely drifts by several meters), and the pins on the map are only approximately the places they name. None of that matters much, because a person closes the gap. You get within ten or twenty meters, glance around, spot the sign or building number for your destination, and walk the last stretch on your own. Close enough is good enough, because the final interpretation happens in your head. But what happens when people start asking AI assistants to navigate the physical world for them, through glasses, earbuds, phones, and eventually robots? The map stops being a picture a human interprets and becomes data a machine acts on directly. A machine can glance around, but not cheaply: giving it eyes means cameras, imagery databases, and the compute to interpret what they see. The AI assistant guiding you might not have the resources to compensate for a pin that’s sitting out in the street instead of at the front door of a business. It acts on what the data says, and it can deliver a wrong answer with complete confidence, or no answer at all. Two things have to be true for it to work: the device has to know precisely where you are, including which way you’re facing, and the map data has to say precisely where everything else is. Zephr’s story with Overture runs through both halves of that problem. ## Teaching a language model where it is Zephr, which Gorman co-founded with CTO Pramukta Rao, took its first run at the language half of the problem through the [Generative AI for Geospatial Challenge](https://aws.amazon.com/blogs/publicsector/announcing-the-generative-ai-for-geospatial-challenge-which-awards-1m-in-aws-credits-to-innovators/), launched in 2024 by Amazon Web Services and the [Taylor Geospatial Institute](https://taylorgeospatial.org/). The premise sounds simple: get a large language model to understand geography. [LLMs, though, are famously bad at it](https://overturemaps.org/blog/2026/open-spatial-location-grounding-for-ai/). They’re trained on text, not coordinate systems; studies asking models to recall city coordinates have measured errors in the hundreds of kilometers, and experiments with generated routes and travel directions find them confidently inventing details. As Gorman puts it, spatial reasoning requires a continuously updating sense of where you are and which way you’re pointed, and a language model has neither. Zephr’s answer was to split the job: the phone’s own sensors establish where the user is and which way they’re facing, in real time, and the model works in language. The GNSS chip is the satellite positioning hardware most people know as GPS. The motion sensors are the accelerometer, gyroscope, and compass that track how the phone moves, tilts, and turns. Before launching Zephr, Gorman had been at Snap, trying to build a crowdsourced 3D map and struggling with “really bad GPS on smartphones in cities.” At the time, the positioning engines were “closed proprietary black boxes sitting on chips in your phone,” Gorman says. But then Google began exposing raw measurements from its chips and he thought, “maybe we can crack open the black box and do something more interesting with it.” Zephr’s founding work was software that sharpens the data from the phone’s hardware to sub-meter accuracy and full orientation. For the challenge, Zephr created a system that converts surrounding spatial data into relationships a language model can reason over. This place is on your left, that road is behind you, the bakery is 200 meters northeast. [By mid-2025](https://medium.com/@zephr.xyz/how-spatially-aware-can-an-llm-be-79f72b5842a2), Zephr had a working AI assistant running on an ordinary phone that could answer quickly and correctly: “is the Empire Restaurant to my right or my left?” and “where’s a good burger nearby?” ## The data problem To describe the world, Zephr needed rich place data (points of interest, or POIs) everywhere. Commercial datasets had the coverage, but their licenses and pricing schemes were a blocker for the scrappy startup. Overture’s openly licensed places data seemed like the answer: Zephr could cache it, correct it, and ship products built on it, without any kind of data contract, using readily accessible open source tools. Work that had once belonged to companies with proprietary data and hefty infrastructure was now within a startup’s reach. When that startup improves the data, the improvement lands in the shared dataset, for Meta and Microsoft as much as for the next small team. Overture’s places coverage was good but the accuracy of the POIs needed work. Zephr’s method is deliberately cheap: it doesn’t analyze a camera feed or match photos against some vast image database. It knows where you’re standing and which way the phone is pointed, draws a line from you toward the storefront, and asks the place data what sits along it. That’s what makes it affordable, and that’s why the answer is only as good as the place data. Point at a real cafe whose coordinates sit out in the street or in the building next door, and the app comes back with the wrong place, or nothing at all. POIs drift for mundane reasons. Places data, in Overture as in most POI datasets, mostly gets its coordinates from geocoders: software that estimates where along a road an address falls. Across millions of records and many data providers, those estimates scatter. Satellite imagery can’t fix it: a satellite sees roofs and the shape of a building’s footprint, not storefront signs. And POIs are the hardest map layer to keep correct anyway: places change constantly as businesses open, close, and change hands. When Zephr analyzed three Colorado cities, roughly 45% of Overture POIs sat outside any building footprint entirely, and about 19% landed in the wrong building. ## Couldn’t have been done alone In September 2025, Gorman brought a proposal to the Overture community: build a shared pipeline that uses street-level imagery to put places pins where the businesses actually are. The proposal worked its way through Overture’s [Map Data Working Group](https://overturemaps.org/about/working-groups/) and was formalized in early 2026 as the [Places Imagery Task Force](https://overturemaps.org/blog/2026/improving-overture-places-geolocation-and-coverage-with-zephr-mapillary-and-the-places-imagery-task-force/). Gorman has seen the inside of plenty of corporate structures. Three of his startups have been acquired, by Esri, DigitalGlobe, and Snap, and he’s collaborated with mapping teams at Meta. Overture, he says, is different. The group is “open and accepting of a small company coming in, and having an idea, and being able to actually execute on it. Really anybody off the street can come in with a reasonable idea and bring it to fruition, which doesn’t happen in most consortiums.” Overture is also the kind of organization Gorman had spent years thinking about, one built around what economists call a club good: shared map data that members jointly fund and maintain, because everyone needs it and no one can justify building it alone. What made the Places Imagery Task Force viable is that Meta and Zephr each arrived holding part of the answer. Meta’s [Mapillary](https://blog.mapillary.com/update/2026/05/21/giving-ai-a-sense-of-place-with-mapillary) team had built a model, running on Meta’s GPU clusters, that scans street-level imagery, detects storefront signs, and hands over clean image cutouts matched to the GERS IDs for Overture’s places data. Zephr’s relocalization method takes it from there: when the same storefront sign shows up in several photos taken from known positions, geometry can recover where it actually sits: trace a ray from each camera back toward the sign, mark where the rays cross, and snap that point to the nearest visible building facade. The pin ends up anchored to the face of a real building. Every correction is keyed to [GERS](https://docs.overturemaps.org/gers/), Overture’s Global Entity Reference System of stable, persistent IDs. Because the relocated place, the building it sits in, and the roads that bound it all carry GERS IDs, a correction can do more than move a pin; it can tie a place to its building and the building to its street. That’s the ambition behind keying everything to shared IDs: a map that behaves less like a stack of separate layers and more like one connected thing, with improvements that any member can pick up. Making these improvements open took deliberate choices. Mapillary’s street-level imagery carries a share-alike license, but Meta grants Overture and OpenStreetMap an exception that extends to derivative work, which is what lets detections drawn from that imagery flow into Overture’s permissively licensed data instead of being fenced in by the license they came from. Zephr made a similar choice: the corrections it produces are released fully permissively, with no proprietary version held back. The bet they’re making is that the durable business isn’t owning the corrected POIs, it’s keeping them fresh. A corrected POI is a fact that decays as businesses open, close, and change hands, so giving it away doesn’t give away the value; the recurring need for current data is the commercial case. ## The results so far On the Zephr side, the task force runs on a small team wearing every hat. Rao worked out the core relocalization method. Nate Ricklin, who Gorman met at DigitalGlobe in 2016, leads engineering and built the pipeline that makes it scale. Scott Nelson handles ground truth, walking streets with survey-grade RTK equipment to record where signs and doorways actually stand. Ovi Voda tunes the pipeline as edge cases surface, and Gorman builds tools to clean messy inputs, sorting out which places belong to which buildings and which sign goes with which business, and writes it all up. On the Meta side, Jake Wasserman and Christopher Beddow support the Mapillary sign-detection pipeline. The team validated the approach across three deliberately varied Colorado test areas: Louisville’s small suburban downtown, Boulder’s Pearl Street corridor, and downtown Denver with its taller buildings and denser urban canyons. Measured against Nelson’s survey baselines, median position error for relocated places fell from roughly five to eight meters down to around two to three meters in Louisville, Boulder, and Denver. The pattern tracks urban morphology: better sightlines and smaller buildings mean bigger gains in accuracy. It also points to where the data and models need to improve next. And once a place is anchored accurately, the same imagery keeps giving. It surfaces places missing from Overture entirely, and it can catch turnover: when the Dunkin’ at a known address becomes a Krispy Kreme, that’s a verifiable, image-backed signal that the place has changed. The ground truthing itself points toward something anyone can participate in. Gorman describes how an AI assistant can prompt a human to help: “I might have my agent navigating you to 12Degree Brewing. When you arrive, it’s gonna say, hey, can you grab a picture? Are the hours posted on the wall? Can you take a picture of the menu?” Because the system knows the phone’s position and pose, and knows it just routed you there, the resulting image is verified on arrival. “Any additional information I can extract from that image, the street address, the sign, the entrance, the hours, the menu, we know that’s good data. And it doesn’t need to be the business owner. It can be anybody. The fact that we mapped it to reality allows us to verify it.” ## Seeing it live The practical output is visible in what Zephr has shipped. Their [Places Grounding](https://zephr.xyz/places-grounding) service is built on Overture data and exposed through MCP, the open protocol that lets AI assistants call external tools. It’s live across the 49 largest metro areas in the United States, drawing on tens of millions of places. Nearly ten million of those places are enhanced with entrances, facades, and visual descriptions that are specific enough to recognize on sight: not just “coffee shop” but the red awning over the patio, the rough-hewn stone facade. At a recent demo session for the Overture community, Gorman asked his AI assistant for a route to a brewery in Boulder. It navigated by landmarks the way a local would: past the barbershop on your left, Patagonia on your right, then turn at Walnut. When Gorman spun the view around and asked again, the first instruction was to turn around. The assistant Gorman demoed runs on the phone itself, not in the cloud. It has to know exactly where you are, moment to moment, and running locally means your position data never leaves the device. And there’s no round trip to a server, so the answers come back at walking speed. Before the Overture Maps Summit earlier this year, Gorman and his team mapped Florence ahead of a talk so they could navigate the city with their app running on Overture data. “I love doing stuff like that,” Gorman says. ## The long way to Pearl Street Gorman didn’t arrive at this work through a single straight line. He holds master’s degree in Geography from the University of Florida and a doctorate in Public Policy and Complexity Science from George Mason University, where his research on complex networks and geospatial infrastructure was unusual enough to be briefed to Congress and included in the President’s Daily Brief during the Arab Spring. Since then, each of his startups has begun the same way: with a research question he couldn’t put down. At GeoIQ, the question was whether mapping and spatial analysis could be pried out of expert hands, so that anyone could layer data on a map and find the patterns his network research had trained him to see. Esri acquired it. At Timbr.io, it was whether the messy, one-off pipelines of data science could be made repeatable and shareable, so that one analyst’s work could become anyone’s. DigitalGlobe acquired that one. At Pixel8earth, the question got more ambitious: could ordinary cameras, crowdsourced at scale, build a 3D map of the globe precise to tens of centimeters, a map made for machines to see rather than for people to read? Snap thought so, acquired the company, and Gorman spent a year and a half rebuilding Snap’s basemap stack before the pull of a new question became too strong to resist. The questions change; the people mostly don’t. Gorman and Rao have worked together since 2005, when Rao joined that first George Mason spinout. He’s been a cofounder of every company since. Roughly two thirds of Zephr traces back to a grad school cohort at UC San Diego (also members of the same climbing gym), and the rest were picked up along the way. “It’s definitely one degree of separation for everybody,” Gorman says. Ask him where the questions come from and he’ll tell you his team is full of “recovering academics.” They post papers in Slack and argue about them, watching for the point where something on the research side meets a gap in the market. “The running joke is that we try to find a barely solvable problem,” says Gorman. By now the rhythm is recognizable: start something small and research-shaped, prove it works, then find a way to scale it. Gorman’s answer has never been an enormous funding round; it’s been a partner big enough to carry the work. Esri, DigitalGlobe, and Snap have each played that role. This time the scaling organization is the Overture Maps Foundation and its open data community. Soon enough, navigating Boulder or London or Tokyo or Lagos won’t involve a screen at all. You’ll ask, your assistant will answer and guide you, and the map underneath will finally know where the front doors are. Not a map you look at. A map you move through. ## Join the Overture community The Places Imagery Task Force is open to any Overture member who wants to help improve the accuracy of Overture’s places data. You can help by contributing street-level, aerial, or satellite imagery, lending engineering time, or helping test the pipeline in a new city or country. Visit our [website](https://overturemaps.org/) to learn more about our work, and [become a member](https://overturemaps.org/become-a-member/). Interested in using Overture data? Check out the [Explorer tool](https://docs.overturemaps.org/blog/2024/07/24/explore-site/), a no-code option for inspecting Overture data and metadata, the [data guides](https://docs.overturemaps.org/guides/) for an overview of our six data themes, and the “[Getting Data](https://docs.overturemaps.org/getting-data/)” section of our documentation. If you’re a [startup using Overture data](https://overturemaps.org/blog/2026/how-10-ai-startups-are-grounding-ai-in-the-real-world-with-overture/), we’d love to hear from you. Reach out to us at community@overturemaps.org. Stay connected: follow us on [LinkedIn](https://www.linkedin.com/showcase/overture-maps-foundation/), [X](https://x.com/OvertureMaps), and [Bluesky](https://bsky.app/profile/overturemaps.bsky.social) for the latest updates, and [sign up](https://bit.ly/3Tfi79o) for Overture’s monthly newsletter.