Feyn

Feyn is a model development partner for teams that want to run a specialist model trained on their own data instead of renting a general purpose one. The engagement is structured in four stages: a review of the current system and agreement on the metrics that define quality; improvements to context and plumbing in the existing stack, which also produce cleaner training samples; training a model fit to the task, covering both the common path and the rare cases that are costly to get wrong; and a production feedback loop so new cases keep improving the model. Customers own the resulting weights. It is aimed at product teams whose AI feature is already live and whose quality ceiling is set by a borrowed model they cannot change. Alongside the paid work, the company maintains open source libraries under the MIT license that anyone can use: an ingestion library for cleaning, chunking, embedding and storing data; a chunking engine built on SIMD delimiter search, lookup tables and zero copy bindings; a unified embeddings client with retries, usage tracking and cost monitoring; and a benchmark for measuring chunking quality across several domains. No pricing is published. There is no self serve signup, trial or plan list, and the only route in is a contact form, so cost is agreed per engagement. The company is based in San Francisco and backed by Y Combinator, and it still operates under its earlier corporate name in site footers and repository ownership, so buyers will encounter the former brand in legal documents and on the open source projects.