TLDC

Market position

Positioned in AI infrastructure and evaluation tooling. Secondary coverage names Pi Labs and Haize Labs as competitors, along with internally built lab tooling, and characterizes the company as early-stage with early production usage among named AI companies.

The company positions its tooling as evaluation-as-infrastructure rather than a results viewer: structured, versioned, reusable eval specifications that can be authored collaboratively across engineering, product and legal, and then reused directly as reinforcement learning reward signals. For its medical model, it argues that general-purpose LLMs trade off groundedness and non-sycophancy for instruction-following on tasks such as coding, and that a domain-specific reasoning model can reach higher clinical benchmark scores at lower serving cost than much larger frontier models.

Technology

Evals-as-code on a YAML schema with versioning across model checkpoints, AI-generated diffs, aligned graders and fine-grained rubrics, plus export of eval specs as reward datasets for GRPO-style reinforcement learning and reinforcement fine-tuning. The company also post-trains large language models, releasing Kos-1 Lite, a reasoning model for clinical use evaluated on OpenAI's HealthBench, a benchmark built with 262 physicians across 60 countries using open-ended conversations.

Go-to-market

Direct engagement with applied AI and frontier model teams, including early access to doteval, help building evaluation datasets, and support for teams adopting GRPO or reinforcement fine-tuning; inbound via a founders contact address. Model releases are announced through the company's research blog with an interactive demo.

Teams building, fine-tuning or operating LLMs at scale, including frontier AI labs, product organizations choosing between models, AI startups building models for regulated industries, and research and open-source teams; secondary coverage cites mid-market enterprises and applied AI teams, with reported users including Perplexity, Diode and Cubic.

Geography

Headquartered in San Francisco, California.

History

The company was founded in 2025 in San Francisco by Daanish Khazi, Gavin Bains and Joseph Besgen and joined a 2025 Y Combinator batch (listed as Spring 2025 by YC, referenced as S25 elsewhere), with Diana Hu as primary partner. It launched publicly as The LLM Data Company with the doteval eval workspace and worked with frontier AI teams on benchmarking complex model tasks. The entity is also listed under the name Paper Instruments, described as post-training LLMs to accelerate knowledge work with open-source frontier models. In March 2026 it announced Kos-1 Lite, a medical reasoning model.

Risks & controversies

Secondary coverage flags early-stage go-to-market execution, heterogeneous customer requirements across domains, competition from evaluation tooling built internally at large labs, and the tension between flexibility and simplicity in product design.