Atlas document intelligence
A contract review pipeline that reads 200-page agreements, extracts obligations and flags deviations from a firm playbook — with citations.
The problem
Atlas's associates spent the first two days of every engagement reading contracts to find the same twenty clauses. It was expensive, tedious and inconsistent between reviewers.
An earlier vendor attempt had failed because it produced confident answers with no way to verify them.
What we did
Every extraction cites the exact clause it came from, so a reviewer verifies in seconds rather than trusting a summary.
We built an evaluation harness first, against 300 contracts the firm had already reviewed by hand. Nothing shipped until it beat the human baseline on recall and matched it on precision.
Ambiguous cases are surfaced as questions rather than answered confidently, which turned out to be the feature the partners cared most about.
The outcome
First-pass review dropped from roughly sixteen hours to under two. Recall on the target clause set sits at 96%, verified quarterly against a held-out set.
Crucially, associates trust it, because every claim carries a citation they can click.
We had been burned by an AI vendor before. Bolteryx built the measurement first and refused to ship until it beat our own reviewers. Every extraction cites its source, so we can verify in seconds.
Have something that needs to exist?
Send the brief. You get a scoped response with pricing, a delivery plan and the names of the people who would build it — within two working days.