About Recension

We built this for the person in legal who has to read a vendor's contract again. What arrives is a master agreement as a PDF, a second PDF about what the subscription covers, and older versions on request. That is simply how contracts get sent, and nobody is at fault for it. But it means the one question that matters, which clause moved since last time, has no answer, so someone hunts for an old copy and diffs it in Word.

Recension is not an app. It is a licensed set of vendor contract documents. Every night we fetch the legal pages a vendor publishes and check whether anything moved. If nothing did, the night ends there. If something did, we keep the file exactly as we found it and cut the text into clauses. Every version is dated and hashed, every clause gets a link you can paste into a ticket, and it all lands nightly in a bucket you own.

Fetching is the easy half. The hard half is telling where one clause ends, and matching this version to the last one. A vendor renumbers a schedule and every path shifts. A PDF is set in two columns, so reading order is not what your eye sees. Two models cover what numbering cannot: a classifier that reads layout instead of numbers, and an embedding index that proposes which earlier clause a new one matches. They propose, a person decides. Below a set confidence we publish no diff at all, because telling a lawyer that clause 9.2 is unchanged when it moved is worse than telling them nothing yet.

The work comes in bursts. Most nights are cheap, because we only re-read what changed. Onboarding is not, and neither is the day a vendor republishes its whole legal estate, or the day we move to a new embedding model and re-embed everything so yesterday's citation still resolves tomorrow. So the bill has three lines. Two are on AWS in eu-west-1, in Ireland, where the product runs. The larger is GPU hours: the layout classifier, the embedding model and the similarity scoring run on EC2 GPU instances we control, raised for an onboarding or a republished estate and dropped again on a quiet night, and from Q2 2027 on reserved capacity with on-demand instances kept beside it for the bursts. The other only ever grows: every fetched file, every version, the clause tree, the embeddings and the manifest hash chain sit in S3, because a citation taken years ago has to keep resolving. The third is training, on Azure Machine Learning in Microsoft's North Europe region, also in Ireland: the classifier and the embedding model are retrained there on the labeled clause pairs, which come from public documents only, and the re-embedding backfill for a new model runs there too. A watchlist names live deals, so the models stay ours, no part of inference reaches a hosted model API, and Amazon and Microsoft sit on our subprocessor register as infrastructure only.

Recension Data Ltd is registered in Ireland with its office in Galway. It is three people and a review desk, and the review desk is the point: every clause pair a reviewer decides is written back as a labeled example, and that corpus, drawn only from documents vendors published in public, is the one thing here that compounds. The price is one published figure per watched vendor, with a ten-vendor floor, because below it there is nothing for the models to be better at than a person with two PDFs open.