- AI Systems
- Prototype delivered
- Client work, anonymised (a consulting firm and its clients)
An AI tool for complex government contract work
A tool that helps reshape complex procurement documents from one contract structure into another, using several specialized AI experts working together.
delivered and under evaluation by the client against their own solicitations, in a closely regulated field. Interface withheld at the client's request.
A look inside
Interface not shown
Client engagement: the interface is not shown. The build is described on this page; a walkthrough is available under NDA.
The work
The problem
Reshaping complex government contract documents between formats is slow, specialized, unforgiving work where accuracy and defensibility matter a great deal.
What we built
A working tool that reads an uploaded solicitation and drafts the replacement document from it. Seven AI "advisors" each read the source against a procurement corpus and return recommendations in their own area; a person accepts, edits or skips each one; only then does the tool write the document, section by section.
The result
A delivered prototype proving AI can be trusted with careful, high-stakes document work: proof of operating in strict, exacting settings, not just with startups.
In more detail
What goes in, and what comes out
A procurement document written one way goes in. A draft of the same work written the other way comes out, in six sections, ready for a person to take over. The tool never files anything and never sends anything; it produces a draft and stops.
Seven advisors, not one model
The work splits into areas a specialist would recognise, and each gets its own advisor rather than one prompt trying to hold everything at once.
- Each advisor reads the uploaded document against a corpus of 291 procurement documents.
- Each returns recommendations in its own area, as a list, with the passage behind it.
- Nothing an advisor says reaches the draft until a person has accepted or edited it.
- Skipping one is a first-class action: a recommendation nobody wanted leaves no trace in the document.
The human is in the loop because the stakes are real
This is work that gets scrutinised by government reviewers, and a confident sentence that cannot be traced is worse than no sentence. The design assumes the tool is wrong sometimes and makes that cheap to catch.
- The person approves before drafting starts, not after it finishes.
- Grounding is checked rather than asserted: a recommendation has to trace to a document in the corpus.
- The tool is scoped to draft and stop. It augments the acquisition professional and never replaces one.
Discipline is the feature
An early version turned a nine-page source into an eighteen-page document, which is the failure mode everybody expects from these tools and the reason people distrust them. Length is now held under test: the same source produces roughly two and a half pages. The regression is locked, so it cannot quietly come back.
Evidence at a glance
- seven grounded advisors, each in its own area
- every recommendation approved, edited or skipped by a person
- a 291-document corpus behind the grounding
- a nine-page source becomes a two-and-a-half-page document, not an eighteen-page one