We teach machines how experts think.
The future of AI won’t be trained on more data, it will be trained on better thinking.
Get data
Explore research
Backed by angels from
Powering every frontier AI research lab
Problem
AI researchers and enterprises are hitting walls with suboptimal data solutions.
Today’s models can generate answers. But they struggle with real work. Because real work isn’t just outputs. It’s decisions, tradeoffs, and context. That knowledge doesn’t live on the internet — it lives inside experts.
Expertise has never been captured. Until now.
The most valuable knowledge isn’t written down. It exists in how professionals think — not just answers, but reasoning, decisions, tradeoffs, and context. We work with domain experts to capture that thinking, then structure it into training data models can learn from.
Our solution
We turn real-world work into training data.
AfterQuery is an applied research lab curating data solutions for frontier foundation model development. Models trained on outputs plateau. Models trained on reasoning improve. We build datasets that reflect how experts actually solve problems — step by step, decision by decision.
Our data includes:
Supervised Fine-Tuning (SFT)
High-quality prompt–response pairs and chain-of-thought reasoning traces — teaching models how to behave across complex tasks.
Reinforcement Learning + Rubrics
Expert-designed prompts with grading frameworks for reasoning and code generation — turning subjective judgment into scalable reward signals.
Agent Environments (API / MCP)
Custom environments across APIs, tools, and services — enabling training and evaluation of agents in real workflows.
Computer Use Trajectories
Human-demonstrated interactions across browser and desktop environments — teaching models to navigate and operate software end-to-end.
Research
Our approach starts with research: where exactly do models break down in real professional contexts? Why do these failure modes exist? We take a proactive stance — every domain has its own failure patterns.
More research
How We Improved Terminal-Bench 2.0 Scores by Over 5x Using Tinker and Harbor
How expert-curated trajectories and tooling lifted Terminal-Bench 2.0 scores more than 5x — and what it says about training agents.
Blog
Mar 31, 2026
Human expertise, reimagined
Capturing how experts think — turning real-world decisions, judgment, and workflows into training data models can learn from.
Blog
Apr 9, 2026
Solving the Last Mile Problem in Partnership with The Raine Group
Encoding domain-specific excellence into forms machines can learn — so agents think and execute like real-world experts.
Blog
Apr 28, 2026