Latent
Latent provides a runtime hallucination detection platform that reads a model's internal state as it generates each token, flagging made-up answers before customers see them. It integrates with vLLM for open-weight models and uses its own reader model for closed APIs, offering 103x faster checks than judge models at no measurable latency. The platform includes a review queue with annotated explanations and can hold, replace, or cut flagged responses.
- Artificial Intelligence
- Developer Tools
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
Large language models frequently generate plausible-sounding but factually incorrect answers, and traditional validation approaches—such as using a second judge model—are slow, expensive, and require sending every answer outside the production environment for review. This leaves teams unable to catch hallucinations in real time before customers encounter them.
Solution
Latent reads the internal state of a model as it writes each token, detecting signals associated with fabrication and holding suspicious answers before they reach customers. The platform integrates directly into vLLM, the open-source inference engine, as a plugin that adds no measurable latency during generation. For closed-weight API models, Latent applies its own reader model to audit outputs. Flagged answers are sent to a review queue where human reviewers can release or confirm failures, with every decision logged in an audit trail for compliance and quality control.
Target Audience
Engineering and AI teams at enterprises and startups that serve LLM outputs directly to customers in production, particularly those in regulated industries requiring audit trails for model outputs, such as finance, healthcare, and customer support.
Features
- In-graph capture within vLLM that reads per-token internal states during generation, supporting early stopping when hallucination risk is detected
- Calibrated scoring on each customer's traffic, adapting to the specific model, domain, and prompt distribution
- Annotated flag explanations showing the specific sentence, driving words, and the most influential source passage, plus failure type (e.g., unsupported number, wrong date, invented name)
- Prometheus endpoint and OpenTelemetry trace IDs for integration with existing monitoring, guardrails, and tracing stacks
- Labelled flagged answers that can be exported for fine-tuning and model improvement
- Self-hosted deployment where no prompts, answers, activations, or weights are sent to Latent's servers
- Works with any model; open-weight models on vLLM are read directly, while closed-weight API models use Latent's own reader model
- SDK supports OpenAI, Anthropic, Gemini, and Ollama clients with fail-open behavior and zero dependencies