⚠️
This demo contains entirely synthetic constituent data. No real names, addresses, phone numbers, or casework details. The synthetic data portrays a fictitious Maryland Senate office ("Senator John Doe") with realistic volume, geographic spread, and issue distribution.
Congressional bill data is real and public, ingested from Congress.gov.
Reference Architecture Demonstration

From Volume to Voice

AI Infrastructure for a U.S. Senate Office — built, demonstrated, deployable.

A working demo of a Senate office's own AI: two independent services running locally — an Answer Engine for plain-English questions over 69,000 real Congressional bills alongside 60,000 mock casework cases and 145,000 mock phone-call logs, and a Learning Engine that continuously teaches it from public sources. All inference local. No cloud.

✓ REAL — Congress.gov bill data, live ⚠ MOCK — All constituent data is synthetic
Launch Demo → 📄 Briefing PDF or read on for context

A large-state Senate office is not a communications operation — it is a signal-processing organization that also does communications. Tens of thousands of constituent contacts per week. Dozens of bills per legislative day. Press inquiries, casework, social mentions, meeting requests arriving faster than any 60-person staff can absorb. Most of this data is read once, then disappears.

This demo shows what changes when the office runs its own AI infrastructure — on its own hardware, inside the Senate enclave, with constituent PII that never leaves the building.

What's in this demo

69,398
Real Congressional bills
50,565
Bills with summaries indexed
60,000
Synthetic casework cases
145,520
Synthetic phone-call logs
107,857
Generated PDF records
158,120
bge-m3 vector chunks

What you'll be able to do

Microservice Architecture

Senate Intel is not a single application. It's a set of independent services, each with its own job, each independently deployable, each replaceable without touching the others. Two are running today. More are in the build plan.

The Answer Engine

Live

Plain-English in. Answers, briefings, and intersections out. Classifies each question, routes to SQL or semantic search or intersection, and renders the result.

  • FastAPI on port 8000
  • pgvector + HNSW index for semantic search
  • bge-m3 embeddings · Qwen 2.5 14B on GB10 for NL→SQL
  • React UI at /app/

The Learning Engine

Live

Watches public data. Teaches the Answer Engine new questions, autonomously. Validated patterns flow into the search index so the demo gets measurably smarter over time.

  • FastAPI on port 6900 · separate service
  • Five pollers: Congress.gov bills + floor votes, govinfo hearings, news, Bluesky
  • Auto-validates every generated SQL pair before indexing
  • systemd timer · every 2 hours · UI at /trainer/

On the Roadmap

Supporting services: a shared Vector Index (pgvector + ChromaDB) feeds both engines; an MCP Admin Layer provides programmatic infrastructure control for office deployments; a Postfix mail gateway (planned) will handle outbound alerts and magic-link authentication.

How the data gets here

Two data sources, two different stories:

Bills — real, from Congress.gov

Initial corpus ingested directly from the Congress.gov public data API. Bill text, sponsors, cosponsors, action history, policy areas, and committee assignments. Each bill summary is chunked, embedded with the bge-m3 model (1024-dim vectors), and stored in pgvector with HNSW index for semantic search.

Daily auto-refresh planned. Currently the corpus represents Congresses 117 and 118 as of the initial ingest.

Constituent data — entirely synthetic

Generated to reflect realistic volume, MD geographic spread, and casework complexity for a large Maryland office. Every name, address, phone number, and case detail is fictitious. Each generated document (outcome letter, incoming letter, email, phone screen, walk-in form, web-form submission) is rendered as a real PDF and stored on disk — the same way scanned office records would exist in a real deployment.

Generated 2026-05-10 / 11. Watermarked "MOCK DATA — FOR TESTING ONLY" on every PDF.

The hardware behind it

A commodity Linux file server (PostgreSQL, pgvector, FastAPI, the ingest pipelines, the React UI) plus a Dell Pro Max with GB10 for AI inference — a desk-side AI workstation built on the NVIDIA Grace Blackwell architecture.

Dell Pro Max with GB10

Chip: NVIDIA GB10 Grace Blackwell Superchip
Memory: 128 GB LPDDR5x unified
Bandwidth: 273 GB/s
Compute: 1 petaflop FP4
Models supported: up to 200B parameters local
OS: NVIDIA DGX OS (Ubuntu)

Three local models loaded today: bge-m3 for embeddings, Qwen 2.5 14B for classification and NL→SQL, Llama 3.3 70B (planned) for drafting and long-form briefings. Total inference stack ~50 GB; 78 GB headroom on the unit.

Open-source throughout. No cloud, no vendor lock-in, no per-seat licensing.

Ready to explore?
Demo is read-only. Clicking, querying, and drilling in are all safe.
📄 Download Briefing (PDF) Launch Demo →