{"id":127060,"title":"Axiomat – Agents that cannot hallucinate","tagline":"Formally verified data queries","body":"Hey YC 👋! We’re [Ethan](https://www.linkedin.com/in/ethan-kharitonov/), [Jack](https://www.linkedin.com/in/jacknaimer/) and [Elad](https://www.linkedin.com/in/elad-perlman/), founders of [Axiomat](https://www.axiomat.com/).\n\n## TL;DR\n\nWe’re building formally verified **AI that cannot hallucinate. Ever.** We’re starting with queries over structured data. Axiomat lets your customers and employees ask AI about their data, even when a wrong answer is unacceptable.\n\n<https://youtu.be/GkoRgSpmd_c>\n\n## The problem\n\n**AI can make mistakes**\n\nAI is ridiculously smart, but even world-class experts can make mistakes. When a wrong answer is unacceptable, the possibility of error forces you to check your agent’s work, restrict what your agents can do, or, in the worst cases, keep AI out altogether.\n\nThis tiny disclaimer does a lot of damage:\n\n![](https://bookface.ycombinator.com/media/?type=post&id=127060&key=user_uploads/4028907/fbb9236c-5770-4313-9cb0-1cccbb520f69)\n\n**People want to use AI anyway**\n\nDashboards are dead. Your users don’t want to learn your interface, and their needs won’t always fit the pages you’ve built. Agents let them ask for what they need without waiting for your team to build another screen. But if the answers “might be wrong” and need to be double-checked, they’re back to the work the agent was supposed to handle.\n\n## The solution\n\nAxiomat gives AI agents a **mathematical guardrail** against all hallucination. We use your business rules to turn English questions into mathematical definitions, then formally verify in Lean that the queries agents make match those definitions. **If we can’t prove it, the query doesn’t run.** The agent retries until it has a proven answer or tells you it couldn’t find one.\n\nBefore it answers, Axiomat shows you how it understood your question. It translates the mathematical definition back into unambiguous English, so you see the exact interpretation the proof covers (and this translation is done with code, not AI!). You never have to wonder whether Axiomat answered a different question than the one you asked.\n\nYou can use Axiomat to launch an assistant you’ve held back or expand one you’ve had to restrict. Those same verified queries can supply the information an agent needs to manage finances or run a business, making this a first step toward trusting AI with those larger tasks.\n\nImagine what we could do with AI if we could actually rely on it.\n\n## The ask\n\nIf correctness is keeping your AI product unreleased or your agent restricted, email [founders@axiomat.com](mailto:founders@axiomat.com) to discuss using Axiomat for the data-querying use case you can’t safely offer today.","slug":"X3M-axiomat-agents-that-cannot-hallucinate","created_at":"2026-11-09T23:00:00.243Z","updated_at":"2026-11-18T02:19:29.221Z","total_vote_count":13,"url":"https://www.ycombinator.com/launches/X3M-axiomat-agents-that-cannot-hallucinate","share_image_url":"//bookface-static.ycombinator.com/assets/ycdc/yc-og-image-c440a0ad1dacfb86eeeb343717479cc54d256614449b4ef719977a0a451f8bc8.png","company":{"id":37104,"name":"Axiomat","slug":"axiomat","url":"https://www.axiomat.com","logo":"https://bookface-images.s3.amazonaws.com/small_logos/73e5297e01ab55a8cd3d54127a27671c6d5e6934.png","batch":"Fall 2026","industry":"B2B","tags":["SaaS","Big Data","AI"],"search_path":"https://bookface.ycombinator.com/company/37104"}}