ymarkov

full report is here https://voygr.tech/blog/does-sf-talk-to-ai/

VOYGR (YC W26) | Founding Product Engineer | San Francisco | ONSITE | Full-time | $150K-$200K + 0.5%-2.0% equity

I'm Yarik, cofounder and CTO. Before this I spent 20 years in tech with 10+ years running infra, ML and search teams at Google, Meta and Apple.

We love maps and build location/businesses APIs for AI apps and agents. If ChatGPT tells you a restaurant is open and it actually closed in March, that's the problem we fix. We check whether a business still exists, whether its hours and details are right, and enrich it with what's being written about it on the web (news, articles, events). Customers get all of it through one API and we solve it for them with our models which use all available sources and ai voice agents which can confirm the information.

You'd be one of the first engineers. Some weeks that's search and ranking, some weeks it's data pipelines or evals, some weeks it's the API or something a customer asked for yesterday. You'll have a big say in what gets built, not just how.

Stack: Python, FastAPI, Postgres/PostGIS, DynamoDB, Redis, AWS (EKS), our own ML models plus LLM APIs (OpenAI, Gemini, Anthropic).

What I'm looking for: - 2+ years shipping production systems, or fewer if your background is unusual and you can show what you've built - real experience with at least one of: search/IR, large-scale data processing, ML systems, API design - you already use Claude Code, Cursor or similar as part of how you work - geo/POI data experience is a plus, not required

Fair warning: we are a small team, fast pace, and there could be nights and weekends. If that's not where you are in life right now, totally fine, this just isn't a good fit.

Process: a short note on why you, a 30 min call with me and then Vlad, a technical deep dive, then a paid trial project.

Apply here: https://www.workatastartup.com/jobs/93490 or email me directly at [email protected]

We are mostly targeting US right now. Businesses always appreciate new leads - e.g. doctors, movers, handymen. Personal agents are coming and they don't have an api to communicate with most of the US businesses yet.

this could definitely be one of the use-cases

Good question - we actually did a benchmark recently and even with maps/search grounding it is pretty suboptimal from a end user's perspective https://news.ycombinator.com/item?id=47366423 For instance, we found this: "The pattern extends to navigation: Gemini’s own Maps grounding makes it worse at giving directions"

Second, Google Maps API has pretty strict licensing terms - you have to delete data within 30 days, have to attribute it properly. There are reasons why OpenAI and other companies are not using their APIs

That is a very interesting insight and thanks for sharing it. We realize that some countries might be having poor maps data, and that example is very helpful - we will take a look

Thanks - we will keep building!

There is a a lot of data publicly available on the web, it can be analyzed with the derived signals stored and then re-checked periodically.

Completely agree - we have been in constant communication with them

Fair point on the repo - cli is in the plans and coming. We've been focused on the API

thanks, looking forward to that!

No, we are not - it is against Google's the terms of service.

open_closed_status is an indicator showing if there are sufficient evidences of the closure. If place doesn't exist, then there are likely no evidence of anything related to it, including closure. It could be confusing, but we will iterate on that.

It doesn't fit into our plans of building this

Fair point - one doesn't instantaneously dethrone Google three months after starting. Here are some benefits of our API. Google's api is recall focused as any search - it shows you all possible places even if that one you are looking for doesn't exists. It is your job to find out if any of the returned results match the one you were looking for. Also, we have more convenient terms of use - you don't have to delete data after 30 days and don't have to attribute data when displaying.

We have models which take all of this into account when producing the verdict. For enterprise clients we emit a calibrated confidence score. With public api we decided to start simpler. Also, we are not using Google data. I’m not a lawyer, but doing that for any maps-related company is simply against Google’s terms of use

That is our vision of where we want to be. There is a lot of information about the places on the public web which you analyze and cross-reference. And we started to solve this problem with validation API which can tell you if a business or point of interest exists at current location.

Appreciate thinking about it, but I think there's some misunderstanding - we don't track people or movements. VOYGR validates places - for instance, we are able to answer a question if this business still open?

It means that there is sufficient evidence found online that this business or poi is located at that address

Appreciate sharing this project - democratizing this data is indeed a very important step. Interesting that you settled on Haiku - did you have a chance to check flash-2.5-lite or gpt-5-nano performance?

Indeed, we've been told about that. Hopefully, it won't be a defining moment for us as a company :)

It is supposed to work if you even don't include unit number or a house number is a bit off. We analyze other signals too, so if the address is a bit off, the API is still supposed to mark a place as existing

We are not providing opening times yet - we just check if place is permanently closed or not. But it is in the works under our experimental enrichment API (which is not yet open to public)

Google also doesn't tell you if a place exists - it just returns the list of possible places which it thinks could be relevant. We have instructions defined for agents to onboard https://github.com/voygr-tech/dev-tools

We are using judges with LLMs and web grounding plus manual grading. We recently did a benchmark on the LLM quality across major AI providers - we plan to open source it soon and will probably open source our API quality check benchmark too https://news.ycombinator.com/item?id=47366423

Thanks! Was it truly down? I have checked and I don't see any disruptions

Thanks! Our initial API works as follows - you provide POI/business name and its address and we are telling you if it exists or not. So if you are looking to check if the plaza is existing, you just need to provide its supposed address. If it is a business within plaza, then an address of that business is required

It is on par with Google Maps API, but Google gives you more data. Our terms of service are more flexible - for instance we don't require attribution and deleting our data past 30 days. And we are actively working on adding more info to our APIs