A federal contractor's morning starts with a stack of solicitations too tall to read, and the winning line is always buried on page forty. Capture teams do not lose to weaker firms. They lose to unread pages. Beltway reads the whole stack, scores every notice like a capture veteran, and backs every sentence with a citation that jumps to the line that proves it.
This is part three of our series on the AI work we have shipped this year. Beltway runs against Ironvale Federal Systems, a demonstration contractor we created so the scoring has a real capability profile to score against. Every screen in the product is live, and every number in it comes from the engine's own precomputed intelligence.
The Work
Government contracting has a trust problem with AI: a summary without a source is worthless at bid time, because nobody takes an unsourced sentence into a bid decision. So the brief was not build an AI that summarizes. It was build an engine whose every claim can be checked in one click.
The result is a capture board a business developer can actually defend in a bid meeting. Read the Beltway case study or open the live engine and click a citation yourself.
What We Built
1. The reading
At build time, Claude reads all 32 solicitations in the library, 185 requirement lines in total, and scores each notice from 0 to 100 across six weighted factors: past performance, NAICS eligibility, set-aside status, agency familiarity, and more.
2. The citations
Every capture brief is assembled from claims that cite the exact requirement line they answer. Click the citation and the source document scrolls to the line and highlights it.
3. The pipeline
Scored opportunities flow onto a capture board. Drag a card from Identify to Qualify to Pursue to Bid and the dashboard reweights the forecast as it moves.
4. The economics
All of the intelligence is computed once, at build time. The live site makes zero AI calls, which is why the total AI bill for the entire library came to seventy-three cents.
What Worked
The unusual decision was refusing runtime AI entirely. Precomputing everything meant every answer could be audited before it shipped, and the running product could be fast, cheap, and incapable of hallucinating on a live screen.
solicitations read and scored against the firm's profile
requirement lines parsed, scored, and citable
total AI cost for the entire intelligence library
The citation mechanic is the whole product in miniature: nothing in a brief has to be taken on faith, because the sentence and its proof travel together.
What It Means for Government Contracting
Capture teams are drowning in exactly the kind of reading AI is best at, and blocked by exactly the kind of trust AI is worst at earning. The industry does not need chattier summaries. It needs verifiable ones.
Beltway's pattern, score against a real capability profile and cite every claim to its source line, is how AI earns a seat in a bid decision. The same pattern fits any compliance-heavy reading problem: contracts, policies, audits.
What this means for you: if your team reads regulated documents for a living, insist on citations before intelligence. An AI that shows its sources is a colleague; one that does not is a liability.
Where the AI Does the Lifting
Claude does the heavy reading once, at build time: parsing requirement lines, weighing six fit factors, and writing briefs in the register of a capture veteran. The product then serves that intelligence statically, so the AI's judgment is in the product without the AI being on the wire.
An unsourced AI summary
- - Reads fast, cannot be verified
- - One hallucinated line poisons trust
- - Useless in an actual bid meeting
- - Costs accrue on every page load
A cited capture brief
- - Every claim linked to its requirement line
- - One click jumps to the proof
- - Defensible in front of a bid board
- - Computed once, served for pennies
What It Does for the Business
For a contractor, the value is a morning that starts with a ranked board instead of a stack: which notices fit, why they fit, and what the forecast looks like if you pursue them.
For Agentify AI, Beltway is the proof that trust is a design decision. The engine's architecture, expensive thinking once, cheap serving forever, is a pattern we now reach for whenever a client's data does not change by the minute.
Final Takeaway: Citations Are the Product
Beltway wins trust the only way software can in a regulated industry: by showing its work. The score is an opinion; the citation is a fact; the product never asks you to confuse the two.
If AI is going to read for your business, make it read the way your best analyst does, with the source open and the line highlighted.
Open the live engine, pick a capture brief, and click one citation. That click is the whole thesis.
