Research

Data quality makes all the difference.

We're driven by the conviction that model performance is fundamentally bounded by training data quality. Through expert collaboration, rigorous curation methodologies, and deep domain expertise, we research datasets that power tomorrow's models.

FinanceQA

A testing suite that evaluates LLMs' performance on complex numerical financial analysis tasks that mirror real-world investment work.

Read paper

Abstract background

IDE-Bench

A comprehensive framework for evaluating AI IDE agents on real-world software engineering tasks through an IDE-native tool interface.

Read paper

Abstract background

Human expertise, reimagined

Spencer M.

·

Apr 9, 2026

Read blog

How AfterQuery Expert Data Drives Model Performance on τ²-bench

Arya F.
Arya F.

Michael E.

Spencer M.

Arya F.

·

Apr 9, 2026

Read blog

How We Improved Terminal-Bench 2.0 Scores by Over 5x Using Tinker and Harbor

Spencer M.

Michael E.

Carlos G.

·

Apr 9, 2026

Read blog

Solving the Last Mile Problem in Partnership with The Raine Group

Carlos G.

Sam J

·

Apr 9, 2026

Read blog

IDE Bench: Evaluating Large Language Models as IDE Agents on Real-World Software Engineering Tasks

Spencer M.

Jeff Y.

Tiana C.

·

Apr 9, 2026

Read paper

Market-Bench: Evaluating LLMs on Introductory Quantitative Trading

Abhay S.

Sam J.

Spencer M.

·

Apr 9, 2026

Read paper

App-Bench: Evaluating Coding Agents on Generating Economically Useful Web-Apps

Andrew Z.

Sam J.

Spencer M.

·

Apr 9, 2026

Read paper

The AfterQuery Thesis

Spencer M.

·

Apr 9, 2026

Read blog

UI-Bench: A Benchmark for Evaluating User Interface Understanding

Sam J.

Agustin G.

Spencer M.

·

Apr 9, 2026

Read paper

Core Research Areas

Computer Use

Multimodal

AI Safety

Data Quality

Evaluation

1

Computer Use

We've created training data and reinforcement learning environments that teach AI agents to navigate real software workflows end-to-end, capturing judgment calls and edge cases that only experienced practitioners recognize.

1

Computer Use

We've created training data and reinforcement learning environments that teach AI agents to navigate real software workflows end-to-end, capturing judgment calls and edge cases that only experienced practitioners recognize.

2

Multimodal

We've produced richly annotated datasets spanning text, images, code, and structured documents, enabling models to reason across modalities the way professionals do when reading a scan, interpreting a chart, or synthesizing information from multiple sources.

3

AI Safety & Security

We've designed adversarial datasets and evaluation environments that systematically expose where models fail in high-stakes professional contexts, stress-testing reliability, alignment, and robustness before these systems are trusted with consequential decisions.

4

Data Quality & Curation

We've developed software-first pipelines for filtering, validating, and scoring training data at scale, ensuring the datasets that power frontier models meet the quality bar that professional-grade AI demands.

5

Model Evaluation

We've created evaluation frameworks grounded in real professional workflows, measuring what standard benchmarks miss: whether a model can structure a deal, reason through a diagnosis, or complete the kind of work that experts actually do every day.

Core Research Areas

Core Research Areas

1

Computer Use

We've created training data and reinforcement learning environments that teach AI agents to navigate real software workflows end-to-end, capturing judgment calls and edge cases that only experienced practitioners recognize.

Computer Use

Multimodal

AI Safety & Security

Data Quality & Curation

Model Evaluation