Reprompt (YC W24)

?? I've been working on Scout, a location agent that roams, analyzes, and can enrich location data with street-level vision. Scout has an extensible environment to freely explore. Today is an integration with Google Maps and an open-vocabulary detection model and now we're integrating other street-level imagery providing. Joining Reprompt (YC W24) has been a challenging and rewarding experience to take on geospatial problems. Scout is a massive leap forward to solve many problems related to delivery, rideshare, POI, and other highly valuable tasks that take humans minutes if not hours to complete. Today, Scout can essentially, ?? Enrich information about a location with a combination of tools like object detection and captioning. ?? Decide combinations of tools to solve problems and assign them to more specialized workflows or agents. ?? Completely control an integrated UI in real-time. ?? Do tons of quick math to get more relevant views derived from a view of interest (the zoom effect). ?? Extend with other tools and functions (Kepler, Mapbox) ?? Return JSON about target attributes given a series of addresses. If enough interest, I'll share a more in-depth article about its inner workings and how it (somehow) was built without any LLM frameworks. We're also looking for talent to find creative ways--much like this agent--to solve difficult geospatial problems.

Meet Scout ?? ??

?? Innovation is all about making the right decisions at the right time. In this conversation with Reprompt (YC W24), Foursquare CTO Vikram Gundeti explains the thinking behind our latest product releases—how we’re pushing the boundaries of location tech, solving real-world challenges, and shaping the future of geospatial data. ??? Jump to the comments for the link to the full video ?? #Foursquare #Innovation #LocationTech #AI

Vikram Gundeti discusses how humans & AI are fixing place data with Reprompt at Foursquare. Watch the entire video from the link below in the comments.

Foursquare OS Places vs. Overture Maps: Why Choose When You Can Have Both? ?? Foursquare OS Places – 100M+ POIs with rich metadata like phone numbers, websites, and operational hours ?? Overture Maps – 40M+ POIs with broad foundational coverage built by Microsoft, Meta, and others At Reprompt, we take the best of both worlds: ? Clean, enrich, and merge Overture’s foundational data with Foursquare’s detailed metadata ? Fix outdated records, validate addresses, and enhance missing attributes ? Ensure production-ready, actionable data—without the manual cleanup headaches With POI data rapidly becoming a commodity, the real edge comes from curated, accurate, and enriched datasets. Read our blog from the link in the comments below ??

Humans??? Agents: Vikram Gundeti explains how Foursquare's vision involves both Humans and Agents collaborating to create the most accurate place data possible. Beside’s the community there are three workforces contributing to Foursquare data: 1. Human workforces in a call center browsing the web and calling businesses and using their local knowledge to update and add places. 2. On-the-ground apps: Users provide real-time corrections and updates directly from their devices. 3. Agents like Reprompt handling tasks at massive scale continually learning from human feedback to match human-level quality. Why are agents like Reprompt so valuable? - They scale tasks that would be impossible for a single person or even a large team. - They reduce costs by automating repetitive work. - They monitor signals—like social media chatter—to see if places are opening, closing, or changing their hours in real time - impossible for a human workforce to do.

Reprompt is the first AI Location Agent used by the FSQ Places Engine. ChatGPT's location data is using Foursquare's place data to ensure high-quality data. The backstory of our partnership: - In November, Foursquare open-sourced its entire place dataset. - We worked with them to develop an LLM-powered Placemaker Agent - Our agent connects directly to their Places Engine and works alongside Foursquare's human Placemakers to complete tasks 200x faster and at 97% accuracy compared to human verification. Learn more about the partnership from the link in the comments.

The combination of reasoning, logic, and access to external information that are all connected to a Generative AI model invokes the concept of an agent. Click the link in the comments to read Google’s white paper on how foundational building blocks of Generative AI agents, their compositions, and effective ways to implement them in the form of cognitive architectures.

DeepSeek R1 is shaking up the AI landscape – matching OpenAI’s o1 reasoning model, making it 50x more affordable. That’s a big deal for NVIDIA and OpenAI, but what does inexpensive intelligence mean for location data? ?? Real-time enrichment at scale: Reasoning models can synthesize multiple data sources – analyzing and extracting insights that previously required human effort. Keeping place data updated is now cheaper and faster. ?? Human-level reasoning: Tasks that once required manual review – custom categorization, lead qualification, and place-based vision analysis – are now possible at scale with exceptional accuracy. ?? Vision reasoning is coming: While o1 and R1 aren’t perfect at vision yet, rapid advances in inference-time compute are pushing multimodal reasoning forward, making complex image and video analysis more effective. At Reprompt, we’re ahead of the curve. Our Location Agent leverages vision and reasoning AI to generate 40+ custom attributes from images, video, web, and social sources. Tap the link in the comments below to improve your operational efficiency.

Foursquare CTO Vikram Gundeti sat down with Reprompt (YC W24) CEO Lukas Martinelli to discuss the brand new Placemaker Agent integration, how we improve our POI dataset with our new Places Engine, and more. ?? Watch the full discussion: https://lnkd.in/ed728kfU Jump to the comments for more about the Foursquare Places Engine.??