We build the layer that lets enterprises run open models on their GPUs with confidence.
Traction Layer AI is a secure LLM inference company based in San Ramon, California. We build a supported, long-term distribution of the SGLang inference engine, secure inference tiers that run inside the customer's own cloud account, and provenance products that prove what a model was given. Everything we ship runs where the customer's data lives.
Why we exist
Open-weight models reached parity with proprietary ones for most enterprise tasks faster than the infrastructure around them matured. The engines that serve them are superb research software and difficult enterprise software: frequent releases, fixes that only go forward, a security surface built for a trusted network, and nobody accountable when something breaks at scale. We spent years running SGLang, vLLM, TensorRT-LLM and Dynamo on NVIDIA hardware for organizations that answer to regulators, measuring them and hardening them. Our products close the gap between what upstream ships and what an enterprise can approve.
How we work
Upstream first.
We do not fork. The scheduler handlers the tuner relies on, our security fixes and our backports go to the SGLang project wherever upstream will take them; the patch set is public; the API is the one upstream documents.
Measured, not marketed.
Every performance number we publish comes from an open-source harness you can run from a clean checkout. If you cannot reproduce it, we correct it.
Plain about what we hold.
We align our logs and controls to the frameworks your assessors use. Our own SOC 2 and ISO/IEC 42001 are on the roadmap, not in hand.
Adjacent to the ecosystem, not against it.
NVIDIA's software stack, Red Hat's platform and the SGLang maintainers' own work are partners and context for us, not targets.
Engineers answer.
Support and sales conversations begin with people who build the branches and benchmark the engines.
Traction Layer AI aligns its controls and audit records to NIST AI RMF, ISO/IEC 42001 and EU AI Act Article 12. Alignment is not certification. Traction Layer AI does not hold SOC 2 or ISO/IEC 42001 certification today; both are on our certification roadmap.
Leadership and team
Shankaran Sitarama
founder, CEO and CTO
More than twenty years of enterprise technology leadership, including Oracle and Xoriant. Doctoral research at UC Berkeley on predicting LLM inference performance from kernel-level structure; inventor on the patent-pending KPG Predictor. Leads product, architecture, the benchmark program and the relationships with upstream and partners.
Engineering.
Inference engineers who work on SGLang's scheduler internals, carry the branches and backports, maintain the benchmark harness and contribute upstream; security engineers who build the provenance and signing stack.
Go to market.
Direct sales and partner channels through GPU cloud operators, systems integrators and cloud marketplaces.
Partners and programs
AWS Partner.
Secure Inference and the RAG Provenance Platform deploy into customer AWS accounts; Traction Runtime runs on Amazon EKS.
Microsoft Azure Partner.
Microsoft Entra ID integration and deployment on Azure GPU infrastructure.
NVIDIA Inception.
Member of NVIDIA's program for AI startups. Traction Runtime is validated on NVIDIA Hopper, Ada and Blackwell accelerators and designed to run beneath NVIDIA Dynamo.
Red Hat Partner.
Collaborating on software supply-chain security built on Sigstore and Red Hat Trusted Artifact Signer; Traction Runtime packages as an operator for OpenShift and works inside llm-d SGLang pools.
STARLabs.
Research collaboration on inference performance prediction and shape-aware kernel selection, the work behind the KPG Predictor.
Partnerships and programs, not customers.
Careers
A small team working on scheduler internals, GPU systems, security hardening, release engineering and cryptographic provenance, with customers who run inference at scale under real constraints. Open roles are on the careers page; if none fits and you think you belong here, write to contactus@tractionlayer.ai with a link to code you are proud of.