Crustdata (YC F24) – Web Search API for Token-Efficient AI Agents
Hi HN! We’re Abhilash Chowdhary, Chris Pisarski and Manmohit Grewal. We built Crustdata (YC F24). Today we’re launching our web search API for AI agents, which not only returns the most relevant documents from the web but also maps them to the correct entity (person, company or event). Demo video here https://youtu.be/IouWW97hBN8 If you run agents at scale, tokens become a line item. The web data is the worst input: long pages, repeated content, mixed entities, stale claims. The usual web search -> scrape -> summarize + structure forces the agent to spend tokens doing…
In plain words
Crustdata is a web search API designed for AI agents that returns relevant documents while automatically mapping them to the correct entities—people, companies, or events. Built for teams running agents at scale, it reduces token consumption by preprocessing web data upstream, eliminating redundant pages, repeated content, and stale information before agents process it. The service maintains a canonical graph of entities with stable IDs and relationships, continuously indexing the web to connect documents to their correct sources.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
Hi HN! We’re Abhilash Chowdhary, Chris Pisarski and Manmohit Grewal. We built Crustdata (YC F24). Today we’re launching our web search API for AI agents, which not only returns the most relevant documents from the web but also maps them to the correct entity (person, company or event). Demo video here https://youtu.be/IouWW97hBN8 If you run agents at scale, tokens become a line item. The web data is the worst input: long pages, repeated content, mixed entities, stale claims. The usual web search -> scrape -> summarize + structure forces the agent to spend tokens doing janitorial work before it can take action. We’re trying to move that work upstream. We keep a canonical graph (ontology) of people and companies: stable internal IDs, aliases, and relationships. Then we continuously index the web and attach each document to the right entity ID. Example: raw web search for "Stripe pricing changes 2026" returns ~10 results across ~4,000 tokens, mostly redundant. We return 6 deduplicated results in ~1,200 tokens. This is not just about saving tokens. It also matters because the common failure isn’t “search missed something.” It’s “search found something about the wrong entity.” Names collide. Companies rebrand. Domains move. Press releases get syndicated and look like independent sources. If you treat strings as IDs, you eventually attach evidence to the wrong person/company and the agent takes a confident action based on that mistake. Under the hood, we run a continuous pipeline that updates the entity-linked index: discover -> fetch -> extract -> dedupe -> entity resolution -> attach -> index . And we serve you this index via our search API. We didn’t start with web search. We spent ~2 years building verified people + company data from higher-trust sources. That forced us to build identity as a system, not a string. When we tried to bolt on web search and started building our integrated index of documents + people + companies, we ended up with a pile of local fixes: parser tweaks, domain rules, prompt hacks. Each fix helped one case and broke another because identity isn’t local. That’s when we committed to an entity-first index: pay the entity resolution cost once, then reuse it everywhere. If you’re building AI agents for sales, recruiting, or investing that do a lot of web searches for people and companies, we’d love for you to try our web search APIs. https://crustdata.com/demo
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
