Each lawyer, court, and agency produces documents in their own formats on their own schedules: PDFs, Word, Google Docs, scanned images, hearing video, and paper -- so much paper. A single legal document can run to thousands of pages, exist in dozens of versions, and cite sources that have since been changed.
Generic ETL tooling treats these documents as blobs of text. That loses the structure: formatting, citations, effective dates, and other critical metadata that legal work depends on. We are lawyers and technologists (and lawyers who are technologists) who put the work in to understand your unique workflows and data sources, and build pipelines that preserve that structure and character-level accuracy, while allowing you to use the latest tools.
A connector is a customized set of transformations that keep your data in sync
We read the publisher's schema and its exceptions: which fields are reliable, where the schema is violated in practice, how often the source updates, and what breaks when it does. This is where most integration projects fail, so it is where we start.
Documents are split into their real units — titles, chapters, sections, subsections — with identifiers, ancestry, citations, and status flags carried along. Both the clean text and the original markup are kept, so nothing has to be re-derived later.
Classification, labeling, summarization, similarity fingerprints, and embeddings are added as pipeline stages with their own evaluation metrics. Each enrichment is measured against known-good data before it ships, and re-measured when a model or prompt changes.
Keyword indexes, vector indexes, and citation lookups are combined so a query returns the right document whether the user typed the exact statutory phrase or described the idea in their own words. Results are delivered through an API you own.
Scheduled loads, backfills, alerting on source changes, and dashboards that show what came in and what failed. The pipeline is code in your repository, deployed on your infrastructure, documented for your engineers.
We publish the output of several of our legislative pipelines as open datasets on Hugging Face, under CC0. They show some of the work we've done on high value data sets. You can download them, read the schemas, and judge the work for yourself.
The United States Code parsed from USLM 1.0.15 into sections, with citations, ancestry chains, repeal status, and both plain text and XML for each section.
View datasetCongressional bill text from govinfo.gov, Congresses 107–117, normalized into structured sections with headers and identifiers alongside the full document.
View datasetBills joined to their policy area and legislative subjects — training and evaluation data for classification against more than a thousand possible subjects.
View datasetBills paired with their Congressional Research Service summaries, the reference set for evaluating machine-generated legislative analysis.
View datasetCommittee referrals for Congressional bills, the metadata needed to route a bill to the people and processes that will act on it.
View datasetPublic datasets, downloaded thousands of times by researchers, journalists, and civic technologists
huggingface.co/dreamproitWe built Smart Search — a combination of keyword matching and semantic search — so MassTrac users can find relevant bills and follow statutory changes even when the searched keywords don't appear directly.
Read the announcementA legislative tracking company was writing meeting summaries by hand, with different requirements in every jurisdiction and a growing backlog. We built a custom pipeline that processes the transcript, identifies the agenda items discussed, and drafts summaries in each client's required format. The backlog cleared and the volume of hearings they can cover no longer depends on headcount.
Read the full storyTens of thousands of bills are introduced each year, many of them reworked versions of earlier text. Comparing 50,000 bills against 50,000 bills, and 300,000 sections against 300,000 sections, is not a job for full-text search. We fingerprint each document with locality-sensitive hashing, store the fingerprints in Postgres, and match on them. This turns days of comparison into seconds. Results are served through a public API built with FastAPI in collaboration with Demand Progress.
The same corpus supports classification. Our labeling pipeline assigns a policy area and legislative subjects to a bill and reports accuracy, precision, recall, and F1 against the Congressional Research Service's own labels, so the quality of a run is a number rather than an impression.
Read the resultsAs a legislative information and bill tracking company, we struggled with manual summarization of state capitol events such as legislative committee hearings for our state-based service, and for generating meeting minutes for our municipal government line of work. DreamProIT provided us with a custom AI pipeline that not only streamlined our summarization and minutes generating processes. They also ensured the timely delivery of precise summaries, helping us to overcome a significant backlog. The results exceeded our expectations. And our clients are thrilled with the results, too. We highly recommend DreamProIT to anyone seeking efficient and reliable AI solutions.
If your firm needs a source of legal or legislative data connected to the systems your lawyers already use — a research tool, a matter database, a drafting workflow, an internal AI assistant that has to cite real authority — we will scope it, estimate it in time and cost, and build it.
