Epsilab GTM brain

Epsilab builds realistic reinforcement learning environments and evaluations for teams training AI agents on long workflows.

category

Agent reinforcement learning environments

market profile

  • text: Frontier model labs need interactive environments and evaluation tasks that model complex, long-horizon work.
  • text: Providers compete on environment realism, task and reward quality, and coverage of professional workflows.

personas

  • title: Frontier model post-training lead
    pains:
    • Static datasets do not capture interactive tasks with long chains of actions and changing state.
    • Training requires scalable environments and trustworthy task grading.
  • title: Agent evaluation research engineer
    pains:
    • Must distinguish model failures from environment, tooling and grader failures.
    • Needs reproducible tasks that reflect real enterprise software and professional work.
  • title: Applied AI team lead
    pains:
    • Agents can fail across long workflows despite performing well on isolated tasks.
    • Testing directly in production introduces risk and makes repeatable evaluation difficult.

icp

  • text: Frontier model developers investing in reinforcement learning post-training for tool-using agents.
  • text: AI labs or enterprise agent teams evaluating realistic software workflows and multi-agent behavior; enterprise demand is a GTM hypothesis.

message pillars

  • text: Train agents in environments grounded in actual use cases.
  • text: Focus on agent training rather than static demonstrations.
  • text: Address real workflows and extended tasks.

objections

  • objection: How representative are the environments of real work?
    response: Inspect task provenance, environment fidelity and held-out workflow performance in a pilot.
  • objection: Can graders be trusted at long horizons?
    response: Review reward design and failure analysis, including whether errors arise from models, tasks, tools or grading.

competitors

  • name: Turing
    website: https://www.turing.com/frontier-ai/rl-environments
    why: Supplies long-horizon RL environments and evaluations to frontier AI teams.
  • name: Scale AI
    website: https://scale.com/blog/rl-environments
    why: Builds RL environments across enterprise, consumer and specialist use cases.
  • name: Datacurve
    why: Supplies agent training data and RL environments for frontier model development.

keywords

  • RL environments for AI agents
  • long horizon agent training
  • agent simulation environments
  • enterprise workflow reinforcement learning
  • agent evaluation benchmarks
  • multi agent reinforcement learning

voice

  • playful visual identity
  • research-oriented
  • concise

signals

  • text: A model lab is hiring post-training or reinforcement learning engineers for agent capabilities.
  • text: An agent team is building evaluations for multi-step enterprise workflows or looking for realistic simulation data.

hiring

  • Technical Strategic Projects Lead
  • Member of Technical Staff, Research Engineer (Reinforcement Learning)

pricing

model: unknown

motion

sales_led

channels

  • YC directory
  • research and hiring networks

confidence

medium

Founders

  • name: Evan Hu
    title: Founder
    role: Founder
    bio: Co-founder at Epsilab (F26). Previously co-founder and CTO at Shortbread (YC W23). Berkeley EECS '24
  • name: Matthew Olsen
    title: Founder
    role: Founder
    bio: Founder at Epsilab - building RL environments for training agents on real workflows and ultra-long-horizon tasks.
  • name: Fengjiao Peng
    title: Founder
    role: Founder
    bio: Cofounder at Epsilab. Previously co-founded Shortbread (YC W23), generating $500K revenue and 2M episodes sold for AI-created comics. ML Researcher at MIT Media Lab.
  • name: Joanne Wang
    title: Founder
    role: Founder
    bio: Cofounder of Epsilab (F26). 2nd time YC founder - previously built Copper Health (W24).

Back to YC companies