CTGT Builds Interpretability Systems to Make Enterprise AI Deterministic

San Francisco-based AI company CTGT is building software to establish a deterministic policy layer for enterprise artificial intelligence. Founded in 2024 out of research at the University of California San Diego, the six-person seed-stage startup raised $7.2 million in seed funding led by Gradient Ventures, with participation from General Catalyst, Y Combinator, Liquid 2 Ventures, Deepwater Asset Management, and angel investors including Mark Cuban, François Chollet, and Mike Knoop. In October 2026, CTGT released an open-source tool named Concordance to trace why AI agents make inconsistent decisions, as reported by VentureBeat.

CTGT develops software centered around its Policy Engine, an infrastructure layer designed to bring deterministic control, auditability, and safety to generative AI models deployed in high-risk enterprise environments such as financial services and healthcare. Rather than relying on separate judge models or prompt-based guardrails, CTGT uses mechanistic interpretability techniques to interact directly with model weights and activation features. As reported by TechCrunch, the platform exposes internal model representations to detect biased outputs and hallucinations, allowing enterprises to intervene dynamically at inference time and refine models without taking them offline.

The company was co-founded by CEO Cyril Gorlla and CTO Trevor Tuttle. Cyril Gorlla left graduate research at Stanford University at age 23 to launch CTGT; his research on efficient, interpretable AI was presented at ICLR while he served as an Endowed Chair's Fellow at UC San Diego. Trevor Tuttle previously built hyperscale distributed systems for large machine learning workloads at MLsys@UCSD. The founders developed their initial approach at UCSD to address compute bottlenecks and bring mathematically grounded interpretability to production AI systems, as noted by The National Law Review.

To expand its Policy Engine and core platform, CTGT is recruiting full-time engineers and interns in San Francisco. A key entry-level technical role is the Machine Learning Engineer: LLM Interpretability & Systems, offering $175,000 to $250,000 per year onsite in San Francisco. The role operates directly with model weights, activations, and architectures to operationalize mechanistic interpretability research into production-grade, deterministic policy enforcement at inference time.

On the software engineering side, CTGT is hiring a Senior Software Engineer ($175,000 to $250,000/year) to make architectural decisions and write code ensuring the Policy Engine remains correct, fast, and auditable under load while defining internal platform APIs. Concurrently, the Senior Software Engineer, Platform ($175,000 to $250,000/year) owns deployment, configuration, and observability systems connecting model work to production environments, operating at the boundary between backend logic and infrastructure. For user-facing controls, the Senior Software Engineer, Product ($175,000 to $250,000/year) designs interfaces and full-stack services that allow enterprise customers to configure complex AI governance rules.

CTGT is also hiring interns for Summer 2027 at its San Francisco headquarters, paying $8,000 to $12,000 per month. The Research Intern: Interpretability & Reliability (Summer 2027) owns a specific problem in feature extraction, runtime intervention, or bounding model error within the Policy Engine during a 10-to-12-week onsite program. The Software Engineering Intern (Summer 2027) builds and ships end-to-end features across the stack for the Policy Engine platform.

In a presentation on YC Root Access, CEO Cyril Gorlla demonstrated how CTGT identifies specific censorship and bias features within open-weight models like DeepSeek R1 to dynamically adjust activations and produce balanced outputs without modifying underlying weights. On the Money Maze Podcast, Gorlla discussed how non-deterministic models create regulatory and operational risks for banks and asset managers, highlighting the need for internal auditability over prompt-based filtering. Speaking at Imagination in Action, Gorlla noted that enterprises are increasingly moving toward open-weight models on-premises to maintain control over sensitive data and meet regulatory standards for explainability.

Open Roles at CTGT

See all open roles at CTGT