Yarik Markov (@YaMarkov) profile picture

CTO @VoygrTech (YC W26) | Building better maps for agents and LLMs

SF/NY
Joined December 2010
We just released PlaceBench: 150 everyday questions about real places in 121 cities, to see how well AI assistants answer them. We tested six: Dots, Muse, Claude Duet, Instinct, Grok Bot and Ask Maps in Google Maps. Most questions name a day and a time, like which Sanborns near the Ángel in Mexico City is still open at 10:30 pm on Tuesday, or the latest you can leave Kings Park in Perth for the last train to Mandurah. 46 are in the local language. Every answer key comes from the place's own website or the operator's timetable. What we found: - Dots scored highest, 92 out of 100, with no wrong answers. It was also the most consistent: right all three times on 17 of 20 repeated questions. - Muse came second at 83, and was the only assistant fully right on the Perth last train. - Claude Duet was the fastest of the top three, with a median of 43 seconds. - Instinct wrote the best group messages, 98 out of 100. - Grok Bot was second best at planning routes and times, after Dots. - Ask Maps answered fastest by far, in 8 seconds, but gave the most wrong answers, 30 of 150. Closed places still catch assistants out. A café in Santiago closed for good in October. Ask Maps, Grok Bot and Instinct said so straight away. Muse said it was open until 7 pm. Questions, answer keys and every grade: github.com/voygr-tech/placeb… Building an assistant or an agent that works with places? We'll run it on PlaceBench and send you the results: voygr.tech/placebench
3
1
5
549
Our SF Tech Week experiment made the SF Standard. We gave an AI a phone with PlaceCall, and 1 in 3 SF businesses kept talking to it AI assistants like Muse or Instinct are starting to call businesses for people. Most businesses are not ready for that yet sfstandard.com/2026/10/29/ai…
224
Huge respect to Supabase and Railway, perfect for prototypes, gtm and etc. Still, for prod I want fewer layers. Supabase was down ~4h in our YC batch, and ours aws-based just kept working. Railway deploys paused 9+ hrs today
2
1
279
We did a fun benchmark - asked Jev, Astra, Fable, Argon and other models to draw a map of SF from memory and scored it against OpenStreetMap. The best ones got 86% of the city right - they know the coast and parks, not the streets. Reverse geocoding from memory isn't here yet, but maybe one day. github.com/voygr-tech/sf-fro…
1
4
198
Yarik Markov retweeted
VOYGR just launched on @ycombinator 's Launch YC! PlaceCall by VOYGR – Give your agent a voice to call any US business, handle reservations, gather information, get quotes. Handles IVRs & hold. Returns outcome + transcript. First 250 calls free. ycombinator.com/launches/Vnj… voyager.tech/placecall
5
3
2
18
44,023
We did an experiment for the SF Tech Week - gave an AI a phone and asked 1000+ SF businesses one question We wanted to understand how ready are SF businesses for AI. Very soon, every personal assistant like Muse, Instinct, Dots is going to call businesses for us, so we wanted to see what happens when it does. Our AI called businesses across SF - vets, dentists, bars, cafés, hair salons etc. It was introducing itself as an AI on every call, then asked a question which fits the business, like "Are you taking new patients?" for a dentist. We called only during business hours. What was found: - only 1 in 3 who picked up kept talking to it - vets 79%, dentists 61% - bars and dry cleaners 20% - most people did not argue - they listened to the intro and quietly hung up ~5 seconds later - about 1 in 25 already had an AI reception answering the phone Caveats - it is just one city and one week, so small differences between neighbourhoods could be a noise voygr.tech/blog/does-sf-talk…
3
3
214
63% of all PRs in our repo (started in December) were merged in the past 5 weeks - one month shipped more than the previous eight combined! We found 30k of expiring OpenAI credits, decided to tokenmaxxx :) Still hiring a founding product engineer - workatastartup.com/jobs/9349…
3
4
176
Just posted about our new product PlaceCall on Hacker News - an API for agents and apps to call businesses and get things done news.ycombinator.com/item?id…
3
3
112
Yarik Markov retweeted
Turns out there's a hard limit on how many friends you can have on a Saturday night in New York. We rang 50 restaurants and asked each one the biggest table they could seat. A table for 50 was the highest anyone offered. 12 were answered by a human. 25 were answered by a bot, which would not discuss capacity and instead asked, repeatedly, how many people were in our party. We hung up first.
1
2
72
Yarik Markov retweeted
On Friday, our CEO @vladibasque helped judge c0mpiled-11, a hackathon in SF organized by @karrwinn and @transposevc. VOYGR had a dedicated prize to award to one of the teams using the VOYGR API to build something cool. Jiwoo Kim built a goal-driven voice agent, Aria, that helps you achieve your long-term goals, including in the real world. You tell it what you're trying to accomplish, such as running a marathon or renovating your kitchen, and it helps you make that happen. Chatbots, while generally useful, don't actually take things from your plate if they have to do with calling for an appointment or comparing options across contractors. Jiwoo Kim's project is different: he used a new VOYGR API (to be announced soon) to look for places that can support the goal’s advancement. In his demo, Jiwoo Kim had the system look for physical therapy appointments when a marathon prep run isn't going well. The product looked up suitable physical therapy options, then called them on the phone one by one to gather specific info, such as appointments available in the evening hours. Huge congrats, and we look forward to seeing where this project goes!
2
4
307
Yarik Markov retweeted
SF billboards had a monopoly on AI for too long. Last week VOYGR was on display in Times Square
1
1
3
314
Claude asked if it should deploy a fix to prod. Then answered its own question. Then deployed. I wasn't even in the room. We are so cooked
2
4
609
Yarik Markov retweeted
Excited to partner with @VoygrTech! Maps tag places. Voygr understands them. Their API validates whether a business exists and is still open, with 99.62% precision. Multi-source verification catches relocations, rebrands, and closures. Real-time ground truth for agents that interact with the physical world. Now on Orthogonal.
1
2
8
611
NYC has all the signs if you just look around
2
412
At Yc demo day - any surface is a pitch surface for @vladibasco
2
7
739
Yarik Markov retweeted
We have launched on HackerNews today news.ycombinator.com/item?id…
2
1
7
379
LLMs ace bar exams, but even the best gets 1 in 12 local queries wrong. We tested 4 leading LLMs (Claude, GPT, Gemini, Perplexity) on 345 real-world local search prompts (finding restaurants, checking hours, planning routes, booking tables), each run with and without web search (2,415 evaluations). Every recommended place was verified against Google Search and Maps. OpenAI leads (90.7/100), followed by Gemini (86.4), Claude (85.9), and Perplexity (80.4). No single provider wins everything; rankings shift by task type, and even OpenAI, the best performer, recommends a place that doesn't exist, has permanently closed, or is in the wrong neighborhood 8% of the time. View report here: voygr-tech.github.io/llm-loc…
3
310