Patronus AI

Patronus AI sells an evaluation platform for LLM and agent applications — managed evaluators, the Percival trace-debugging agent and open judge models such as Lynx and GLIDER — and, more recently, RL environments and Digital World Models for training and testing AI agents on simulated digital workflows. It targets AI engineering teams and frontier labs.

Patronus AI was founded in 2023 by Anand Kannappan (CEO), who helped develop explainable ML frameworks at Meta Reality Labs, and Rebecca Qian, who led responsible NLP research at Meta AI, and launched with a $3 million seed round led by Lightspeed Venture Partners to evaluate and test LLMs for regulated industries. Its evaluation platform remains on sale: a Python SDK and web app for experiments, logs and traces, plus a managed evaluator library — Glider for fast guardrail and rubric checks, Judge for customizable LLM-as-a-judge scoring, Judge MM for image and audio, and Lynx for hallucinations — covering context relevance, context sufficiency, answer relevance, enterprise PII and toxicity. The company publishes much of its research openly: Lynx, a Llama-3 fine-tune released in 8B and 70B sizes in July 2024, whose 70B version scored 87.4% on the company's HaluBench versus 86.5% for GPT-4o, and GLIDER, a 4B Phi-3.5-mini judge released in December 2024 that grades against user-defined rubrics; the Lynx 70B and GLIDER weights carry a non-commercial CC-BY-NC-4.0 licence. Percival, an agent that reads traces from LangGraph, crewAI, OpenAI Agents SDK, Pydantic AI and smolagents, detects 20+ failure modes and suggests fixes. Since late 2025 the emphasis has shifted toward training-side simulation: Generative Simulators, RL environments that co-generate tasks, world dynamics and reward functions, arrived in December 2025, followed by Digital World Models — language diffusion world models for agent training and evaluation — announced with a $50 million Series B in June 2026 led by Greenfield Partners, bringing total funding to $70 million. Patronus says revenue grew more than 15x over the prior year and that it works with the majority of leading frontier AI labs and hyperscalers, placing it against evaluation vendors such as Arize, Galileo and Braintrust as well as agent-training data suppliers.

Category

Developer Tools

Pricing

Freemium, Usage-based, Subscription, Contact for pricing

Target Market

CTOs, AI/ML Engineers, Enterprise Developers, AI Research Labs

Deployment

Cloud-first, API-based

Founded

2023

Headquarters

San Francisco, USA

• ✓
  Patronus Evaluators

  Managed evaluator library — Glider, Judge, multimodal Judge MM and Lynx — scoring hallucination, context relevance, answer relevance, PII and toxicity without building judge infrastructure.

• ✓
  Lynx hallucination model

  Open-weight Llama-3 fine-tune in 8B and 70B sizes that returns PASS or FAIL with reasoning on whether an answer is faithful to its source documents.

• ✓
  GLIDER judge model

  4B-parameter Phi-3.5-mini fine-tune that grades text, conversations and RAG output against user-defined rubrics, returning reasoning, highlighted key phrases and an integer score.

• ✓
  Percival agent debugger

  Agent that analyses traces from LangGraph, crewAI, OpenAI Agents SDK, Pydantic AI and smolagents, detects 20+ failure modes and suggests fixes, improving as users annotate issues.

• ✓
  Experiments, logs and tracing

  MIT-licensed Python SDK and web platform for tracing functions, running experiments and comparing LLM application versions so quality regressions surface before release.

• ✓
  RL environments (Generative Simulators)

  Adaptive environments that co-generate tasks, world dynamics and reward functions for training agents on long-horizon, real-world-like digital workflows.

• ✓
  Digital World Models

  Language diffusion world models, unveiled June 2026, that scale creation of simulation data to train and evaluate agent actions across complex digital workflows.

• •
  RAG faithfulness gating

  A financial-services RAG assistant runs Lynx or hallucination evaluators on each answer, flagging responses unsupported by the retrieved documents before analysts rely on them.

• •
  Agent failure debugging

  An engineering team sends LangGraph or crewAI traces to Percival, which identifies failure modes across 20+ categories and proposes optimizations to the agent workflow.

• •
  Release regression testing

  Before shipping a new model or prompt, teams run experiments against curated datasets and compare evaluator scores with the previous version to catch quality regressions.

• •
  Custom policy and PII checks

  Compliance and platform teams define rubric-based Glider or Judge evaluators for company policy, style and enterprise PII, then use them as guardrails on production outputs.

• •
  Frontier agent training

  A frontier lab trains long-horizon agents on Patronus RL environments and Digital World Models that simulate software, research, communication and enterprise workflows.

The evaluation platform publishes list pricing: a free Developer tier with no credit card, a $25/month Base plan, and metered evaluator API calls at $10 per 1,000 small-evaluator calls, $20 per 1,000 large-evaluator calls and $10 per 1,000 eval explanations. On-prem or dedicated VPC deployment, SSO, custom data retention and custom evaluator fine-tuning are Enterprise-only and quoted on request. RL environments and Digital World Models have no published pricing.

This page was written from 15 sources, 9 on domains other than patronus.ai.

1. 1.patronus.ai — patronus.aivendor
2. 2.patronus.ai — pricingvendor
3. 3.patronus.ai — percivalvendor
4. 4.patronus.ai — rl environmentsvendor
5. 5.patronus.ai — blogvendor
6. 6.patronus.ai — patronus evaluatorsvendor
7. 7.prnewswire.com — patronus ai raises 50 million series b and unveils first dig
8. 8.techcrunch.com — patronus ai conjures up an llm evaluation tool for regulated
9. 9.venturebeat.com — meet patronus ais lynx the open source bullshit detector out
10. 10.huggingface.co — Llama 3 Patronus Lynx 70B Instruct
11. 11.huggingface.co — glider
12. 12.huggingface.co — PatronusAI
13. 13.github.com — patronus ai
14. 14.pypi.org — patronus
15. 15.hn.algolia.com — search