CTGT: The deterministic layer for frontier intelligence

CTGT

The deterministic layer for frontier intelligence

Fall 2024

Active

Artificial Intelligence

The deterministic layer for frontier intelligence

CTGT is an applied AI research laboratory fundamentally solving the alignment and reliability bottleneck for enterprise AI. For enterprises, especially highly regulated industries, deploying Generative AI is historically a compromise between capability and catastrophic risk. Standard enterprise approaches, such as RAG, fine-tuning, and prompt engineering, operate at the wrong abstraction layer. They are inherently probabilistic, carry massive engineering overhead, and fail to deliver the mathematical certainty required by the Fortune 500. We focus on the science of representation engineering and have productized mechanistic interpretability. By opening the "black box" of neural networks, CTGT has developed a proprietary architecture that intervenes directly at the model's representation layer. We convert complex corporate SOPs, SEC/FINRA regulations, and strict editorial rulebooks into machine-readable "Policy as Code," enforcing deterministic constraints and defensible audit trails without requiring expensive model retraining. The result is a step-function breakthrough in enterprise AI economics and capability. Our fundamental architecture allows organizations to run secure, self-hosted open-source models that mathematically match the reasoning and performance of frontier models. Benchmarks from our enterprise deployments demonstrate a 96.5% prevention of hallucinations, up to a 3.3× accuracy multiplier in complex domain-specific tasks, and an 80-90% reduction in human-in-the-loop manual review. Backed by an $8M seed round from Gradient Ventures (Google), General Catalyst, and Y Combinator, CTGT is currently deployed with Fortune 500 companies, including Tier-1 financial institutions and global media conglomerates, giving them the deterministic control necessary to deploy enterprise AI with zero margin for error.

Active Founders

Cyril Gorlla

Founder

Cyril left his research at Stanford at 23 to found CTGT. His work on efficient and interpretable AI was presented at AI conference ICLR while he was the Endowed Chair's Fellow at the University of California San Diego. He is a Nordson Leadership Scholar and Ivory Bridges Fellow.

Cyril Gorlla

Founder

Trevor Tuttle

Founder

Trevor built hyperscale distributed systems for large machine learning workloads at MLsys@UCSD.

Trevor Tuttle

Founder

Trevor built hyperscale distributed systems for large machine learning workloads at MLsys@UCSD.

The Problem: “Black Box” of AI

The status quo of deep learning involves throwing more compute at models with the expectation that they will achieve true intelligence. The questionable nature of this assumption aside, this will certainly lead to more complex models that we don’t meaningfully understand.

I’ve been obsessed with AI and its potential to change the world since I was a kid. In high school, I took apart old laptops to eke out just a bit more performance to train models, so I’m exceedingly aware how important access to and understanding AI is. (For more background, check out the great piece TechCrunch did on us.)

Enterprise companies are sitting on powerful AI models but can't deploy them due to hallucinations and unpredictable behavior.

Healthcare providers can't risk medical chatbots giving dangerous advice
Financial firms need guarantees their AI won't recommend fraudulent investments
Tech companies need AI that consistently aligns with their brand voice

Our Solution: Principled AI Development

CTGT takes a fundamentally different approach. Rather than black-box approximations, we've developed breakthrough technology that understands how neural networks actually learn.

This allows us to:

Automatically detect and eliminate hallucinations
Train models 10x faster with 10x less compute
Give enterprises foundational control over AI behavior

How it works:

Let’s imagine we ask an LLM about cooking food.

I started lunch by simmering water on the stove. I prepared dough for a homemade pizza dish. Using a knife, I sliced fresh tomatoes and cheese. I spread glue on the bread for tomorrow's breakfast. The dough rose in the oven. I brewed tea while stirring the soup with a spoon. A plate was set, and lunch was nearly complete. Every pot and utensil had its place.

We can determine, for example, that pizza and bread is part of the “food” concept in the model, along with other related things in the output.

https://www.youtube.com/watch?v=WSDro9tEz_w

And we can determine that “glue” shouldn’t be part of this concept.

https://www.youtube.com/watch?v=skW5Z5CPr9o

Our technology ensures “glue” doesn’t appear in queries regarding food, reducing hallucinations by 80-90%.

https://www.youtube.com/watch?v=sH8x9boOtF4

Real Results with Fortune 10 Companies

We're already helping major enterprises deploy AI with confidence:

Removed bias from facial recognition systems in one step
Eliminated investment advice hallucinations for portfolio managers
Enabled fraud detection LLMs with precise error identification

The Team

Founded by UCSD researchers pushing the boundaries of AI:

Cyril Gorlla (CEO): Recipient of the Endowed Chair's Fellowship at UCSD. His work on efficient AI training was invited for presentation at ICLR and his talk at Ai4 was the most attended in its track. He collaborated with Intel on ML telemetry deployed on 8M+ CPUs. He was named one of 12 “2022 Shining Stars” at UCSD and is an Ivory Bridges Fellow and Nordson Leadership Scholar. He was advised by ACM and Amazon Fellow Mikhail Belkin.
Trevor (CTO): Deep learning researcher focused on high powered distributed ML systems under Hao Zhang, a student of Databricks Founder Ion Stoica, at UCSD.

Why Now

Legacy companies and those without access to massive amounts of compute can greatly benefit from specialized, domain-specific AI. Current LLM interpretability methods are extremely inefficient, effectively rendering them inaccessible to the vast majority of companies. There’s a shift from large foundation models to those tailored to each company’s brand and applications. Our technology enables this future by making state-of-the-art AI accessible without massive compute resources.