The Problem
Congressional bills are dense legal documents, and Congress introduces hundreds of them each week. Most people do not have time to read one, let alone keep up. Civic organizations want to tell their communities what is in the bills that affect them, but writing summaries by hand cannot keep pace.
What We Built
A website, built for a civic transparency organization, that does the reading. Every day it checks Congress.gov for new activity, sends each new bill to an AI model to be summarized, and publishes the result on a public page. Bills are sorted by where they stand: just introduced, in committee, on the floor, or sent to the president. Nobody has to press a button for any of this to happen.
How the AI Is Kept on Task
An AI model will write a confident summary of almost anything. The work is in deciding, ahead of time, exactly what it is allowed to do. Three decisions do most of that here.
It works only from the bill
A model has absorbed plenty of news and opinion about well-known legislation. Commentary is not the bill, so the instructions tell it to use the bill's own text and nothing else.
Every summary has the same shape
A few key points for someone with thirty seconds, then a fuller breakdown one click away. A fixed format leaves the model less room to wander.
The source is one click away
Each summary links to the full bill on Congress.gov. A reader who wants to check a point can go straight to the original.
These choices were worked out with the client before launch. The tighter the instructions, the less a system like this depends on someone reading every output.
What It Does
- Collects new bills from Congress.gov every day, automatically
- Writes a summary in two layers: key points first, detail on request
- Sorts bills by where they are in the process, from introduction to the president's desk
- Links every summary to the original bill text
- Shows the client what the AI is costing to run, on an admin page
Screenshots
Common Questions
How do you know the summaries are good?
Quality is handled in stages, sized to what a small organization can support. Before launch we defined what a good summary is and reviewed a first set of real bill summaries with the client, who approved them for public release. The platform also has a reviewer workflow built in, so domain experts can rate summaries on several aspects of quality. It is ready to switch on when reviewers are available, and those ratings will guide further tuning.
What does the AI cost to run?
Less than most people expect. Summarizing a bill costs about a third of a cent ($0.003) under our design constraints, which for the current Congress comes to roughly $50 a year in AI usage. AI models are billed by how much they are used, so cost was treated as part of the design: usage and spend are tracked from the start, and the client can see both on an admin page.
Would this work for other kinds of documents?
Yes. The pattern is general: gather documents from wherever they live, have an AI model summarize them or pull out key details under clear limits, and deliver the result where your team already works. That could be a web page, an email digest, a spreadsheet, or a record in your database, always with a way back to the source. It fits best where the documents or the rules are particular to your organization: reports that must follow one funder's template, applications judged against your own criteria, years of scanned records in a format nobody else uses, or notes that have to be matched to records in another system.
Behind the Build
For technical readers.
The platform runs on Flask with a PostgreSQL database, deployed on Railway with scheduled jobs for daily ingestion from the Congress.gov API, filtered by action type. Summaries are generated through the Claude API under a fixed prompt template, with rate limiting and retry logic for reliable API calls. The feedback data model stores reviewer assessments across multiple quality dimensions, and cost monitoring tracks Claude API usage.