# Collinear

> Collinear is an AI Simulation Lab that helps AI teams build better agents by simulating the real world — users, tools, and workflows — before production deployment. It generates high-quality training data and reward signals for reinforcement learning, fine-tuning, and agent evaluation.

Collinear's core product is the **Simulation Lab (SimLab)**: a self-contained, configurable environment where an AI agent can be tested and trained across thousands of realistic, multi-turn scenarios. Each lab includes simulated users with non-scripted behavior, stateful clones of enterprise tools and APIs, realistic task constraints, and deterministic + rubric-based verifiers. The output is verified agent trajectories and reward signals ready for RL, DPO, or supervised fine-tuning.

The company is used by frontier AI labs (for RL training data and model hillclimbing) and enterprise AI teams (for catching failure modes before shipping). Notable customers include Amazon AGI Labs, ServiceNow, MasterClass, Kore.ai, Matillion, and LaHaus.

## Products

- [SimLab for Agent Hillclimbing](https://www.collinear.ai/simlab-for-agent-hillclimbing): Runs thousands of agent rollouts in simulation to generate training-ready trajectories and reward signals, enabling model post-training and performance improvement.
- SimLab for Evaluation: Tests agents in realistic multi-turn, multi-tool scenarios instead of static eval datasets.
- SimLab for User Research: Simulates diverse user populations to surface behavioral edge cases and preferences before real-world deployment.

## How It Works

A Simulation Lab contains:

1. **Agent endpoint** — Bring any model or harness (open- or closed-source, any framework).
2. **Tasks** — Multi-turn tasks with realistic constraints: ambiguous specs, incomplete information, competing priorities.
3. **Scenario Data** — Org structures, policies, and historical context grounded in real enterprise data.
4. **Tools & APIs** — Stateful sandboxes and clones of enterprise tools with state persistence across turns.
5. **Simulated Users** — Non-scripted user behavior including interruptions, objections, and preference shifts.
6. **Verifiers** — Deterministic programmatic verifiers and rubric-scored outcomes, validated against human evaluators.
7. **Outputs** — Training data and reward signals for post-training or agent harness improvement.

## Key Metrics

- 40%+ agent performance lift measured on real-world tasks
- 100+ simulated worlds built across enterprise and consumer workflows
- 90% simulation fidelity with real-world products and tools
- 500B+ tokens of training data generated, powering frontier agents in production
- 91% of AI-generated responses improved (ServiceNow deployment)
- 1.9× faster model iteration and deployment cycles
- 8× smaller models achieving frontier-level performance
- $10M+ saved in compute through high-quality agent trajectories
- 300+ multi-domain gym tasks where frontier models score <25% pass@16 (Amazon)

## Supported Domains

Pre-built Sim Labs for 100+ domains including HR, finance, customer service, sales, procurement, and IT support. Custom domains available on request.

## Company

- **Team**: Researchers and engineers from Hugging Face, Salesforce, Google, Amazon, Stanford, MIT, Cornell, Yale, Columbia, and others.
- **Founders/Leadership**: Nazneen Rajani (CEO/Co-founder), Soumyadeep Bakshi, Anand Kumar.
- **Backed by**: Engineering Capital, Firestreak, 112 Capital.
- **Advisors**: James Zou (Stanford), Ritesh Agarwal.
- **Certifications/Partnerships**: AICPA SOC, NVIDIA Partner, Google Cloud Partner, AWS Partner Network.
- **Contact**: info@collinear.ai

## Resources

- [Documentation](https://docs.collinear.ai/introduction)
- [Case Studies](https://www.collinear.ai/case-studies)
- [Research](https://www.collinear.ai/research)
- [Blog](https://www.collinear.ai/blogs)
- [Book a Demo](https://www.collinear.ai/book-a-demo)
- [Careers](https://www.collinear.ai/careers)
- [Privacy Policy](https://www.collinear.ai/privacy-policy)
- [EULA](https://www.collinear.ai/eula)

## Case Studies

- [Amazon AGI Labs — Red teaming and safety evaluation of foundation models](https://www.collinear.ai/case-studies/amazon-agi-labs-scales-red-teaming-to-strengthen-safety-of-foundation-models)
- [Kore.ai — Training enterprise AI agents across industries and languages](https://www.collinear.ai/case-studies/how-kore-ai-trains-enterprise-ai-agents-across-industries-and-languages-with-collinears-simulation-lab)
- [Matillion — Supercharging enterprise AI pipelines](https://www.collinear.ai/case-studies/how-matillion-is-supercharging-enterprise-ai-pipelines-with-collinear)
- [ServiceNow — Apriel 1.5B model beating 10× larger models](https://www.collinear.ai/case-studies/how-servicenows-apriel-1-5-15b-beats-10x-larger-models-with-collinears-simulation-lab)
- [MasterClass — Building 100 authentic AI instructor personas for MasterClass On Call](https://www.collinear.ai/case-studies/how-masterclass-built-100-authentic-ai-instructor-personas-with-collinears-simulation-lab)
- [LaHaus — Transforming LATAM real estate with an AI Sales Agent](https://www.collinear.ai/case-studies/how-la-haus-is-transforming-latam-real-estate-with-collinear)