Personal project · 2026
EO Digest retrieves new executive orders from the Federal Register, generates plain-language summaries, validates their citations, and publishes them automatically with links to the original documents.
0
manual steps from publication to summary
4 stages
ingestion, summarization, validation, and publishing in one pipeline
Verification boundary
No public demo or source repository is linked. This case study documents the implemented pipeline, its public data boundary, and the verification performed without presenting concept art as a product screenshot.
Executive orders often run thousands of words and use dense legal language. EO Digest needed to turn each official document into a shorter, readable summary while keeping the original source easy to verify. The main constraints were automating the full publishing flow, keeping summaries grounded in the source, and keeping operating costs practical for a personal project.
Built independently, with early feedback focused on summary readability.
The Federal Register API provides structured publication data and canonical links. Using it avoids depending on another website's markup.
Orders can appear on the Register after they are signed, so the site follows the official publication schedule rather than breaking news.
Each ingested order retains its Federal Register metadata and permanent source link.
Free-form summaries can lose important source details. Requiring citations for each section makes the output easier to verify.
The prompts are more rigid, and a summary may need to be regenerated when validation rejects it.
Validation rejects references to nonexistent sections before publication, and every summary links to its source document.
Content changes only when a new order is published, so it does not need to be generated on every request.
The publishing pipeline must trigger revalidation; if that step fails, a page can remain stale.
Pages are served from static output while each new publication triggers revalidation.
Symptom. Early generated summaries occasionally described effects that the source order did not support. That made source fidelity the central quality problem, not a secondary refinement.
Investigation & alternatives. I considered stitching together shorter chunk summaries, adding a second model as a verifier, and constraining the output format. Chunking lost document-level context, while a second model added cost. I chose a structured schema that requires each summary section to cite the source section it uses, followed by validation that rejects references to sections that do not exist.
Verification & result. In spot checks across a sample of orders, the schema and validation prevented the unsupported claims seen in earlier outputs. Every published summary also retains a link to the original Federal Register document for direct review.
The finished system connects scheduled ingestion, LLM summarization, output validation, a relational data model, and a statically rendered frontend. A new document can move from official publication to a source-linked summary without manual publishing steps.
What I’d change. Quality review is still based on sampling rather than an exhaustive manual check. I would expose more validation detail in the interface and design full-text search into the initial data model instead of adding it later.
Next project
Playlist Builder