Find AI

Find AI was a startup where I served as CTO and founding engineer from 2023-08 to 2024-11. I built the engineering team and the core product: a semantic, multi-criteria search engine over unstructured profiles of people and companies.

People used Find AI by asking questions like, “Find software engineers who went to my university and own a dog.” The system decomposed each question into criteria, generated candidates in parallel, and reviewed the evidence for every result.

I built the search system in Ruby on Pinecone and Elasticsearch. At production scale, it processed 19 million OpenAI requests per week over more than 100 million vectors. I also designed the API, built its TypeScript SDKs, and implemented usage-based billing. Clay was an early enterprise customer and offered Find AI directly inside its product. The company raised $8 million from Felicis Ventures, Daniel Gross, and Nat Friedman.

• Vector similarity was not accuracy. A vector search could understand that a Chihuahua and a golden retriever were related, but it could also confuse one dog breed with another or a former job title with a current one. Find AI got better results by breaking complex questions into smaller criteria, combining retrieval methods, and asking an LLM to judge the evidence.
• The pipeline was agentic, even though the team did not describe it that way. Rather than asking one model for a complete answer, the system decomposed a goal, gathered candidates and evidence in parallel, and made a judgment about each criterion. The simple search box hid a multi-step analysis system.
• Evals were production infrastructure. I designed an LLM search-evaluation harness that covered query decomposition, parallel candidate generation, LLM-as-judge scoring, regression benchmarks, and evals. Another internal tool ran the same production prompt across several models, measured cost, and asked a judge model to score and critique every response. That let the team test whether a model that cost 100 times less was good enough before changing millions of calls.

The lesson I carried forward is that LLMs can be analysis engines alongside writing and chat tools. They can turn unstructured data into useful decisions when retrieval, evaluation, and observability are treated as first-class systems.

In late 2024, the company chose a different direction. I left, and the original Find AI search product and API are no longer active. I wrote about the transition, and what the experience changed for me, in “Lost and found”.

Find AI API demo. This short video shows the customer experience for searching people and companies through the Find AI API. I also wrote a technical tour of the API, including its OpenAPI foundation, generated client libraries, documentation, and usage-based billing.

Behind the scenes of Find AI. In this 2024 talk at Mindstone in New York City, I walk through the semantic-search and model-comparison tools the team used to understand retrieval quality, model behavior, and cost. The accompanying post is “Internal tools of Find AI”.

For a deeper look at production observability, model evaluation, batches, and fine-tuning, watch “How I use data to optimize AI apps”.