Tinfoil
Private AI inference
Tinfoil makes AI workloads secure and provably private: the privacy of an on-prem deployment while running in the cloud.
- Founders
- Tanya Verma, Jules Drean, Sacha Servan-Schreiber
- YC batch
- P25
Company posts
Summaries from the company’s own website.
- Safety Without Compromising on Privacy ↗
Tinfoil says it evaluates models before hosting them and runs safeguard models inside secure enclaves during chat to flag harmful outputs without exposing conversation content. The post argues this approach can preserve privacy while enforcing a narrow set of safety rules, with public, auditable policies and open-source enforcement code.
- Architecting Secure Prompt Caching ↗
Tinfoil says it implemented prompt caching for its Inference API while preserving privacy by using client-side cache secrets and enclave-based routing. The post explains how cache salting, cache-aware routing, and volatile GPU-resident storage are used to reduce compute and latency without letting the operator probe customer cache contents.