Félix Balazard

Today we can finally talk about @enveda's partnership with @AnthropicAI . “Partnership” is a word that covers a lot of sins, so here is what it actually bought us: 1. a Phase 2 program pulled forward by three weeks 2. an IND submission accelerated for another soon-to-be-announced clinical program 3. an Investigator’s Brochure, drafted by @claudeai and signed off by our team after one minor edit, now in front of the FDA We chose to work with Anthropic on development, not discovery. Enveda already finds medicines faster than we can develop them, and I think the industry is heading the same way. The slowest part of development is paperwork, directly or indirectly: batch records, study reports, the Investigator’s Brochure, and the IND that packages it all for the FDA. Discovery gets the headlines. Document review is where the months go. Test 1: manufacturing batch release for a Phase 2 program. Before a batch can go to a trial, someone checks every page of its manufacturing record by hand. Claude read 1,061 scanned, largely handwritten pages from the campaign. Four of our reviewers and outside consultants independently reviewed the same pages, and a Quality adjudicator scored every comment against one question: would an FDA auditor raise it? Claude matched all 11 actionable issues the experts found and added 7 they had missed. Review time fell from 26 hours (typically spread over several weeks) to under 3, an 89% reduction, for about $37 in model cost. On another program, it flagged a climbing impurity that routine review had cleared; it later went out of specification. Test 2: the Investigator’s Brochure for an IND, drafted in four hours instead of weeks. For a new Enveda medicine about to enter human trials, Claude drafted the brochure every trial physician relies on, from a structured layer of facts pulled from our tox, bioanalytical, and PK reports. Our reviewers found one minor unit-conversion error. We fixed it and included that draft in the IND. From the same source layer, Claude drafted the core of the IND itself in four hours, work that takes the best teams a month or more. None of this works unless the model stays factful, so we built a custom harness to force it. It reads every page as an image, keeps a code ledger of which pages it read, and ties every finding to its page and governing document. Across the evaluation, no finding cited anything that was not in the record. The full method is public. Our technical report on the CMC Document Processor covers the architecture, GMP controls, study design, and results, and we open-sourced a companion implementation so you can build your own! Thank you to the incredible @enveda and @AnthropicAI teams for making this possible.