CTGT is an artificial intelligence (AI) safety and governance company based in San Francisco that was established in 2023 and uses interpretable AI models to support the development of reliable enterprise AI. CTGT's technology supports organizations in deploying AI reliably by removing hallucinations, bias, and compliance risks through real-time monitoring and direct surgical interventions in neural networks. As a trusted partner of many Fortune 500 companies, CTGT enables organizations to develop trustable AI solutions up to 500 times faster than other options in industries such as healthcare and finance.
Founded by Cyril Gorlla and Trevor Tuttle, both University of California, San Diego (UCSD) dropouts who were also accepted to participate in the Y Combinator accelerator program.
Raised an oversubscribed $7.2 million seed round led by Gradient Ventures with participation from YC and General Catalyst.
Featured on Y Combinator with 7 employees working together to create safeguards for general artificial intelligence (GenAI).
Proprietary neural network interpretability technology using statistical methods for feature isolation/modification, real-time model monitoring, and efficient training acceleration for LLMs.
Inferred from company descriptions; specific languages/frameworks not publicly detailed
| Service | Cost | Details | Source |
|---|---|---|---|
| Managed Plan | Custom quote (annual billing) | Cloud-hosted platform for AI model evaluation, safety testing, and training optimization. Pricing based on product tier and usage. | Official terms and conditions |
| On-Premises Option | Same annual fee as Managed Plan | Self-hosted deployment for full data control and no data leak concerns. Same pricing as cloud version. | TechCrunch article |
| Feature | CTGT | Gorilla (competitor method) | OpenAI Evals | Anthropic Guardrails |
|---|---|---|---|---|
| Model Safety Testing | Yes (mathematical guarantees) | Partial | Yes | Yes |
| Real-time Monitoring | Yes | No | Partial | Partial |
| On-Premises Deployment | Yes | No | No | No |
| No Additional Training Required | Yes | No | Partial | No |
| Cost Optimization (500x faster) | Yes | No | No | No |
| Active Model Retraining | Yes (no downtime) | No | No | Partial |
| Interpretability Guarantees | Yes | No | Partial | Partial |
| Enterprise Data Control | Full control | Cloud only | Cloud only | Cloud only |
| Starting Price | Custom enterprise | $20+/mo | $20+/mo | Custom |
| Free Tier | N/A (enterprise focus) | Yes | Yes |
CTGT provides mathematically-guaranteed interpretability of the AI system without requiring additional training of judge models, unlike OpenAI's evaluation techniques which do not provide this capability. CTGT has lower blended costs than OpenAI ($0.38 vs. $15).
CTGT would be better suited for enterprise compliance and cost sensitive safety testing; OpenAI would be a better option for rapid prototyping.
In addition to being enterprise-focused, both CTGT and Anthropic can be deployed on-premises and allow for real-time retraining of the AI system. However, CTGT uses significantly less compute resources (is 500 times faster at training) than Anthropic does in their approach to constitutional AI.
CTGT is more flexible in terms of deployment and cost effective for enterprises; Anthropic has a stronger position in the area of ethical AI frameworks.
CTGT claims to have superior performance through its use of mathematical interpretability rather than the judge model based approach used by Gorilla. CTGT also doesn't require computing intensive training of the monitor models to achieve this goal. While both platforms are targeted towards enterprise AI safety, CTGT is attempting to accomplish this goal more efficiently.
Positioned as an efficient alternative to the current state-of-the-art methodologies.
Enterprise-focused platform versus open-source; HF does not offer real-time monitoring, on-premises deployment, or guaranteed interpretability. CTGT is designed for Fortune 500 companies and HF is targeted at developers.
CTGT for production enterprise; HF Evaluate for research and experimentation.
HaluEval identifies fabrication in text
HaluEval exposes bias in facial recognition and decision systems
We have mathematically-guaranteed methods to evaluate your models without needing to retrain them to judge them
TruthfulQA and category specific enterprise benchmarks
| Benchmark | Category | Supported |
|---|---|---|
| HaluEval | Hallucination Detection | Yes |
| TruthfulQA | Truthfulness | Yes |
| Legal Reasoning | Regulated Industries | Yes |
| GSM8K-equivalent | Math Reasoning | Yes |
We can perform surgical interventions on latent features
We can suppress confabulation at the activation level
Flags babbling before user exposure
