Klavis AI GTM brain

Klavis AI supplies reinforcement learning environments and training data to AI labs building capable agents.

category

AI agent training environments

market profile

  • text: Frontier labs need training and evaluations for agents that perform multi-step work, beyond answering prompts.
  • text: Klavis combines stateful environments with task-specific verification and expert-curated data.

personas

  • title: Head of post-training
    pains:
    • Static question-answer datasets do not teach agents to recover through multi-step workflows.
    • Building reliable rewards and sufficient task diversity internally consumes research time.
  • title: Agent evaluation research lead
    pains:
    • Hard to measure progress on evolving codebases and tool use reproducibly.
    • Hand grading long trajectories is expensive and inconsistent.

icp

  • text: Frontier model developers with dedicated reinforcement learning or post-training teams.
  • text: AI labs training coding, scientific research or professional-work agents; exact employee counts are not stated.

message pillars

  • text: Train on persistent, realistic work environments.
  • text: Provide trajectories across multiple demanding domains.
  • text: Make outcomes measurable through verifiers.

objections

  • objection: Will the tasks fit our target capability?
    response: Explore domain-specific tasks and a pilot with the team; verify fit before committing to a large dataset.
  • objection: Can we trust the reward signal?
    response: Assess the programmatic tests and task-specific rubrics on held-out examples.

competitors

  • name: Datacurve
    why: Offers expert-sourced coding data and RL environments for AI labs.
  • name: Idler
    why: Builds RL environments and evaluations from production work.

influencers

  • name: Xiangkai Zeng
    url: https://www.linkedin.com/in/xiangkai-zeng
    why: Klavis founder discussing agent training and integrations.

keywords

  • RL environments for AI agents
  • agent training data
  • long horizon coding benchmark
  • agent reward verifiers
  • tool use evaluation
  • frontier model post training

voice

  • technical
  • research-oriented
  • ambitious and concise

signals

  • text: Hiring researchers for agent post-training or reinforcement learning suggests a need for environments and rewards.
  • text: Publishing long-horizon coding or tool-use evaluations suggests relevant agent capability investment.

stack

  • Docker
  • Python
  • MCP

hiring

  • Research Engineer

pricing

model: unknown

motion

sales_led

channels

  • research publications
  • GitHub
  • X
  • LinkedIn

confidence

medium

Founders

  • name: Xiangkai Zeng
    title: Founder/CEO
    role: CEO
    bio: Co-founder @ Klavis AI. Previously Senior Software Engineer on Google Gemini at Google Deepmind. Co-authored Gemini paper. I have a CS MS from Carnegie Mellon University. I have a Bachelor of Engineering in Computer Science and Technology from Beihang University.
  • name: Zihao Lin
    title: Founder/CTO
    role: CTO
    bio: Co-Founder at Klavis AI 1. Senior Software Engineer at Lyft, tech lead at Lyft recommendation team and Nordstrom data infra team, building product supports millions of users in the last 5 years. 2. CS MS from Northeastern University, Bachelor of Engineering in Computer Science and Technology from Beihang University. 3. The President of the Computer Science Student Council (CSSC) at Beihang University. Research Experience at the NLP Lab and Internship at Baidu.

Back to YC companies