AfterQuery

What AfterQuery Does

AfterQuery builds training data that teaches AI models how domain experts think and work. The company's core thesis is straightforward: today's models can generate answers, but they struggle with real work — the decisions, tradeoffs, and contextual judgment that professionals exercise daily. That knowledge, AfterQuery argues, does not live on the public internet; it lives inside experts and their organizations.

The company encodes this expert thinking into structured training datasets within the broader AI Data Mining space. Instead of starting with agents and working inward, AfterQuery starts with data — understanding each organization's atomic workflows, internal precedents, and decision patterns — and works outward.

Public Validation

AfterQuery's approach has attracted attention from frontier AI labs. NVIDIA named AfterQuery as the only data partner in the Nemotron 3 Ultra technical report, using AfterQuery's Off-The-Shelf Office Agent Training Dataset to improve the 550B-parameter model on GDPval, OpenAI's benchmark covering 1,320 professional tasks across 44 occupations. The training used a technique called PivotRL, which reuses intermediate decision points from AfterQuery trajectories to teach models the judgment steps between decisions.

In a separate effort, AfterQuery researchers built a two-stage post-training pipeline — Tinker and Harbor — that improved openai/gpt-oss-20b from 3.1% to 17.0% on Terminal-Bench 2.0, surpassing Gemini 2.5 Flash. Critically, the fine-tuning used zero overlap with the evaluation set, meaning the improvement reflects genuine generalization rather than memorization of benchmark tasks.

How the Training Works

AfterQuery's fine-tuning produces a distinct behavioral shift that separates it from conventional training approaches. Base models tend to start writing code immediately from assumptions about the environment. AfterQuery-trained models consistently begin by exploring: listing files, reading documentation, and understanding the environment before acting. This exploration-first workflow emerged organically from training on successful trajectories — the team found that reward shaping for exploration produced only performative behavior, not genuine understanding.

Training episodes are capped at 20 turns with no context summarization to prioritize speed, while evaluations use default settings with no turn limit and summarization enabled. This asymmetry between training and evaluation conditions makes the benchmark results more meaningful.

Enterprise Approach

AfterQuery's most distinctive operational pattern is its data-first, forward-deployed model. When partnering with The Raine Group, an investment bank, the AfterQuery team spent three days on-site in the firm's New York offices working alongside bankers to build Raine Search, a semantic search tool for querying the firm's precedent library using natural language. Investment banking, AfterQuery notes, runs on precedents — junior bankers spend roughly half their time adapting off-the-shelf materials because there is not enough time to build every deliverable from scratch.

This same pattern extends to AfterQuery's broader offering: custom datasets, off-the-shelf training data, enterprise AI consulting, and end-to-end implementation, all accessed through direct engagement rather than self-serve platforms.