Paper Instruments: Building AI Tools for Knowledge Work
Paper Instruments builds open-source frontier models and agent-first tools for knowledge work. Formerly known as The LLM Data Company, the startup began with tooling for evaluating LLMs and agents and has expanded into software that helps AI agents work with real-world professional files and workflows.
Visit paperinstruments.com ↗Basic Information
| Field | Detail |
|---|---|
Company Name |
Paper Instruments |
Former Name |
The LLM Data Company |
Category |
AI Infrastructure |
Subcategory |
AI Agents, Open-Source Models, Evals & Knowledge Work |
Founded |
2025 |
Headquarters |
San Francisco, California |
Stage |
Seed / Early Stage |
Reported Funding |
$125,000 Y Combinator seed investment |
Founders |
Daanish Khazi, Gavin Bains & Joseph Besgen |
YC Batch |
Spring 2025 |
Team Size |
Small founding team; public profiles list approximately 3–8 employees |
Website |
paperinstruments.com ↗ |
What Paper Instruments Does
Paper Instruments develops software and models for AI agents performing knowledge work. Its current focus includes open-source frontier models and agent-first tooling for manipulating professional documents while preserving their underlying structure.
Key Products & Research
- Paper Office for agent-safe Word, PowerPoint and Excel editing
- Open-source tools for structured professional documents
- Model and agent evaluation infrastructure
- Doteval workspace for writing and executing evals
- Research into reinforcement learning and post-training
- Frontier model projects including Kos-1 and Kos-1 Lite
Why It Matters
AI agents increasingly need to modify the documents people already use, not simply generate new text. Paper Instruments is targeting this infrastructure layer, where preserving formatting, formulas, relationships and document structure becomes part of agent reliability.
Paper Instruments at Scale
92.5%
Paper Office Benchmark
Paper Instruments reports a 92.5% pass rate across five models and 61 document-editing tasks when its packages are used with the accompanying guidance.
3
Office Packages
Paper Office includes separate Python packages for Word, PowerPoint and Excel workflows.
2026
Research Expansion
Research has expanded from evaluation tooling into frontier model work, reinforcement learning and agent-first software infrastructure.
Product & Research Momentum
Paper Office
Released in September 2026, Paper Office provides Python packages that let agents edit existing Word, PowerPoint and Excel files while attempting to preserve the structures users did not ask them to change.
From Evals to Agent Tools
The company’s original product, Doteval, focused on writing, versioning and executing evaluations for models and agents. Paper Instruments now applies the same emphasis on measurable agent performance to the software layer agents use.
Frontier Model Research
The company has published work on Kos-1, Kos-1 Lite, medical reinforcement learning and DiligenceBench, an evaluation for long-form equity research agents.
Funding Journey
2025
Founded
The LLM Data Company
Founded by Daanish Khazi, Gavin Bains and Joseph Besgen
Spring 2025
Y Combinator
S25
Joined Y Combinator and launched tooling for model and agent evaluations
2025–2026
Evals & RL
Doteval
Developed evaluation workflows, reinforcement-learning datasets and frontier-model research
September 2026
Paper Instruments
Paper Office
Expanded into agent-first tools for structured Word, PowerPoint and Excel editing