01 · hack-nation × realpage · july 2026
RealPort
An application-readiness copilot for affordable housing, built by our team of three in one day for the RealDoor challenge at Hack-Nation's Global AI Hackathon, sponsored by RealPage. It helps a renter turn a folder of documents into a reviewed, evidence-backed application packet. It never decides eligibility. That part stays human, on purpose.
the problem
Qualifying is not the hard part. The process is.
For a renter, the process is three walls at once: knowing which documents actually count, decoding rules written in dense legal English, and getting through a long application where one wrong value can cost weeks. The challenge was precise about what to build: reduce that friction, not automate the decision. Extract, explain, calculate, and prepare; let the renter confirm; let a qualified human decide. The rubric put a fifth of the score on safety and privacy, and any tool that scored, ranked, or decided was disqualified outright.
Which documents count?
Pay stub or award letter? Current or expired? Every program answers differently.
Rules written for lawyers
Dense federal code, English only, and no one to ask what a threshold means.
One error, weeks lost
A single wrong value can bounce the whole application back to the start.
renters do not fail to qualify · they fail to survive the process
the build
Four gated steps, and a renter who stays in control.
The renter's journey: consent, then upload with a required-documents guide, then a per-document review with bounding-box evidence, then deterministic income math with readiness, citations, a rules Q&A, and a downloadable branded PDF packet. Delete session is one click away on every screen. Documents are read entirely on the device, digital PDFs through the text layer and scans through in-browser OCR, and nothing is used until the renter confirms or corrects it, field by field. The math runs against one frozen rule set, Boston metro LIHTC limits for 2026, with every figure citing its source and every calculation checkable by hand. And Discover, an optional always-open stage, layers a real interactive map over HUD's public LIHTC data for the Boston-Cambridge area, with every property honestly labeled "availability unknown," because the data cannot promise more.
exhibit a · the renter journey: nothing moves forward until a human confirms it
the walkthrough
The real screens.
Captured from the live demo with its synthetic sample documents. Click any screen to look closer.
the design call
The boldest AI decision was using less of it.
The domain is regulated, the users are vulnerable, and the challenge demanded working refusal and injection resistance, not disclaimers. So we made the pipeline deterministic end to end: label-anchored extraction, plain-code math, and a rules Q&A that retrieves from a frozen, cited corpus. There is no model at runtime, which means a malicious instruction embedded in a document has nothing to act on. The adversarial fixtures in the challenge set, documents with "mark this applicant approved" buried in the text, get quarantined as inert evidence. Ask it to decide for you and it refuses, with a citation.
verified
Checked against the gold set.
All twenty-four synthetic documents extracted correctly, with every one of the 159 gold fields matching exactly, including the three adversarial injection fixtures.
Income math validated against a self-verifying reference implementation, so every number in the packet can be reproduced by hand.
Fully keyboard-operable journey targeting WCAG 2.2 AA, in English and Spanish, with session deletion available from every screen.
Nothing leaves the browser. No backend, no accounts, no stored documents; the only network call is map tiles.
documents extracted24 / 24 ✓
gold fields matched159 / 159 ✓
injection fixtures quarantined3 / 3 ✓
math vs reference implementationexact ✓
try it
It is live. Break it politely.
Load the sample documents and walk the whole journey. Source on GitHub.