DOGE Factory

Inspiration

A better government app cannot fix a bottleneck written into the rules.

We built DOGE Factory around a simple question: What if AI could help modernize the policy behind a public service—not just its interface?

Housing and healthcare became our starting points because administrative requirements affect both the people seeking assistance and the staff delivering it. Our goal is to identify opportunities for improvement while preserving verification, privacy, oversight, and appeal rights.

What it does

DOGE Factory connects three steps in one workspace:

Audit: Retrieve relevant policy and identify potential barriers, grounded in exact source passages.
Redline: Draft targeted replacement language with explanations, safeguards, and uncertainties.
Mobilize: Carry a reviewed proposal into a stakeholder-engagement workflow with voice-agent support.

Users can explore housing, healthcare, or a broader problem library, review findings, compare proposed changes, and revisit saved analysis runs.

Our golden-path example

HUD guidance requires two consecutive pay stubs when pay stubs are used for Level 4 wage documentation.

DOGE Factory examines that requirement and drafts a proposal exploring equivalent, auditable wage evidence—while preserving verification and tenant protections.

The result is a proposal for expert review, not a claim that the policy has changed or that existing alternatives are unavailable.

How we built it

We built the interface with SvelteKit and TypeScript, the policy-analysis backend with Go, and research and voice-agent services with Python.

HydraDB stores a graph connecting policy documents, passages, jurisdictions, public officials, organizations, and lobbying disclosures.

Vercel AI Gateway powers policy analysis and drafting. The voice workflow integrates Claude, ElevenLabs, and Twilio.

Our data foundation

Our November 14, 2026 database snapshot contains:

Metric	Records
Graph nodes	87,697
Relationships	156,761
Policy documents	3,329
Policy passages	66,112
Actors: officials, lobbyists, and organizations	13,565

These are stored records across three collections, not globally unique entities.

Sources include public agency directories, congressional rosters, and federal lobbying disclosures. Relationships retain supporting sources so users can trace connections back to the evidence.

We also enriched all 529 official records with election metadata:

329 have a next regular seat-election date.
200 are marked not applicable for their appointed or non-elected role.

Election dates describe the seat’s schedule, not a prediction of reelection or confirmation that an official will run.

Challenges we ran into

The hardest challenge was preserving the distinction between evidence and inference.

A lobbying disclosure does not automatically establish influence over a particular policy. Committee membership does not necessarily establish authority to amend an agency rule. A documentation requirement may apply only under specific conditions.

We designed sourced relationships and verification metadata to keep those distinctions visible.

We also worked through:

Inconsistent source formats and naming conventions.
API rate limits during disclosure collection.
Database request-size limits during bulk imports.
Passing structured policy context between analysis and voice services.
Keeping generated recommendations separate from sourced facts.

Accomplishments we’re proud of

We built a source-backed path from a plain-language goal to a reviewable policy proposal.

We’re especially proud of:

A connected policy graph with traceable evidence.
A clean original-versus-proposed policy comparison.
A housing demonstration grounded in an actual HUD passage.
A searchable workspace spanning housing, healthcare, and other policy problems.
A Mobilize workflow that can consume reviewed proposal context.

What we learned

Policy modernization is not simply removing steps. It requires understanding what a requirement protects, then evaluating whether another approach can preserve that protection with less burden.

We also learned that trustworthy AI needs explicit uncertainty and clear boundaries between sourced facts and generated recommendations.

The most useful output is not just a persuasive answer—it is a proposal someone can inspect, challenge, and improve.

What’s next for DOGE Factory

Validate the housing example with a policy expert and pilot partner.
Connect the actor graph to policy-specific target selection.
Expand state-level lobbying disclosure coverage.
Verify current guidance and existing alternatives before recommending changes.
Measure administrative burden using real operational data.
Rehearse a complete, expert-reviewed Audit → Redline → Mobilize demonstration.

Our goal is to help turn policy friction into specific, defensible improvements to public services.