Company
Latent catches the answers a model makes up.
We build the check that reads a language model while it writes and sends the answers it cannot vouch for to your reviewers.
Built by people from
Research published at
What Latent is
Latent reads your model's internal state as it writes and holds the answers it is making up before a customer sees them. It runs inside your environment, as a plugin in the vLLM you already serve and a review service beside it. Nothing is sent to Latent. It works with any model. Closed-weight API models are audited by Latent's own reader model. Open-weight models served on vLLM are read directly, which adds per-token risk and actions inside generation, such as early stop.
Founder
Models leave a trace when they make something up, and I built Latent so the teams who stand behind every answer can catch it before their customers do.
Published LLM research at ACL 2023, EMNLP 2025 and NeurIPS 2025. Previously did research at Stanford and Princeton across reasoning, evaluation and interpretability, and built ML at Robinhood and Amazon. Dual degree in computer science and finance from Penn's M&T program (Wharton).
How to reach us
- Book a call
- latent.cal.com/vedant
- Security review
- The security detail
- Vulnerability reports
- Disclosure policy