What changed
FACT: CFPB finalized revisions to the Section 1071 small-business lending data rule (Federal Register, 2026-05-01), changing institution/transaction coverage, the small-business definition, required data points, and the compliance date. FACT: OpenAI released gpt-oss open-weight models runnable locally via Ollama. FACT: PP-OCRv6 offers 50-language OCR at 1.5M–34.5M parameters, small enough for CPU/self-hosted extraction. HYPOTHESIS: together these make an on-prem extraction pipeline technically cheap for the first time.
Why now
Every covered lender must rework 1071 data collection to the revised data points before the new compliance date (opportunity framing is inference per the source note). The privacy objection to cloud AI tools — sending borrower financials to a third-party API — just lost its technical excuse: local open-weight models plus sub-35M-param OCR run on a commodity server inside the lender's perimeter.
Converging signals
(1) Regulation: revised 1071 rule forces re-tooling of data pipelines [federalregister.gov]. (2) AI capability: gpt-oss local inference removes API cost/data-egress blockers [ollama.com]. (3) AI capability: PP-OCRv6 makes document extraction viable without cloud OCR APIs [huggingface.co]. Genuine three-way convergence; the regulation supplies the deadline, the two capabilities supply the on-prem form factor.
Customer pain
HYPOTHESIS: compliance officers at covered banks, credit unions, CDFIs and fintech/equipment-finance lenders must extract ~20 data points per covered application from heterogeneous loan files and file them accurately, and the revision means work already done against the 2023 rule must be redone. NOTE: demand_evidence array is EMPTY — no complaints, forum pain, or job postings were retrieved to confirm this pain, so it remains inferred from the rule text's existence, and demand/spend scores are held down accordingly.
Who pays
The covered lender (compliance/lending-ops budget). HYPOTHESIS with high prior: regulated FIs already pay for HMDA/fair-lending software, so a budget line exists — but no spend evidence was provided in this input, and the revised rule may have RAISED coverage thresholds, which would shrink the buyer pool toward larger institutions with formal vendor procurement (unverified — must be checked in the rule text).
Solved today
HYPOTHESIS (industry-standard, not evidenced in input): incumbent HMDA/fair-lending suites (Wolters Kluwer, Ncontracts, Asurity RiskExec) are adding 1071 modules; larger lenders use LOS vendor add-ons; smaller lenders plan manual spreadsheet collection or consultants.
Why current solutions are bad
Incumbent modules assume structured LOS data and charge institution-scale prices; manual extraction from PDFs/scans is slow and error-prone; cloud AI extraction tools hit an immediate data-governance wall at banks. The on-prem angle is the only real differentiation here.
Proposed product
A self-hosted appliance (Docker image) that ingests loan-file PDFs, runs PP-OCRv6 + a local gpt-oss model to extract the revised 1071 data points, runs deterministic edit/validity checks against the final rule's requirements, flags low-confidence fields for human review, and exports a submission-ready file. Sold per-institution annual license plus setup.
MVP version
Pipeline for the 10 highest-friction data points on the 3 most common document types (application form, tax return, financials): OCR → extraction → confidence-scored review queue → validation report. Benchmarked accuracy on a public/synthetic loan-file set as the demo asset. 4–8 weeks solo with AI-assisted build; founder can fund the GPU box and test data.
30-day build
Read the final rule's actual compliance dates and coverage tiers (not yet verified from the provided excerpt); build extraction MVP; interview 5–10 compliance officers at small covered lenders and 3 compliance consultants to test the fatal question: will a bank buy this from a solo vendor at all?
60-day build
Pilot with 1–2 friendly small lenders or, more realistically, license the engine to a compliance consultancy that already has bank trust; publish extraction-accuracy benchmark as demand-gen.
90-day revenue plan
Paid pilot ($5–15k) via a consultant partner or a CDFI/equipment-finance lender outside the bank-procurement gauntlet. Honest estimate: first revenue is more likely at 6–9 months than inside 90 days given FI vendor due diligence.
Distribution path
Weakest link. Compliance-consultant partnerships, state banking association vendor directories, HMDA/1071 webinar circuit, LinkedIn compliance-officer niche. No self-serve motion exists for bank compliance software; every sale passes through vendor management review.
Pricing hypothesis
$6k–$20k/yr per institution depending on volume, plus paid pilot/setup. Per-filing pricing (the founder's ELDT model) fits poorly because 1071 filing is annual/batch, not per-transaction.
Technical difficulty
Moderate. OCR+LLM extraction is squarely in founder's wheelhouse; the hard parts are extraction accuracy good enough for regulatory data (errors become the lender's federal filing errors) and maintaining validation logic against rule changes.
Legal / regulatory risk
Meaningful: output errors feed a federal regulatory submission; expect contractual liability pushback, and note 1071 has a history of litigation and political reversals — a further delay or rescission kills the deadline pressure (this history is background knowledge, marked HYPOTHESIS, not from provided sources).
Platform dependency
Low — self-hosted stack of open-weight components is the point. gpt-oss and PP-OCRv6 licenses permit commercial use (verify exact license terms — unproven here).
Founder fit
Mixed. The regulation-compels-filing shape matches his proven ELDT edge (lesson applied, confidence 0.8), and the build is his sweet spot. But the buyer class does NOT match: ELDT customers were small training providers who could swipe a card; 1071 buyers are regulated FIs with vendor due-diligence, SOC2 expectations, and 6–18 month trust cycles — exactly the 'long trust-building play' he avoids. The capital-and-runway lesson (0.9) softens but does not remove this.
Breakout potential
If the on-prem extraction engine works for 1071, the same appliance extends to HMDA, BSA/CDD document processing, and general on-prem loan-file intelligence — a real expansion path, but each step deepens the enterprise-sales problem.
Final recommendation
REVISIT, do not build yet. The convergence is real and the mandate-shaped structure is right, but this fails the founder's channel test today: the only buyers are institutions he can't sell to solo, incumbents own the shelf space, and the system retrieved no demand evidence. Worth 2–4 days of validation (compliance dates, coverage thresholds, 5 compliance-officer conversations, one consultant-partner conversation). If a consultant-channel or small-lender segment with card-swipe-ish buying emerges, re-score; otherwise archive.
Next action
Read the Federal Register final rule to extract the actual compliance dates and coverage thresholds, then cold-message 5 compliance officers at small covered lenders and 2 fair-lending consultants asking how they plan to collect the revised 1071 data points — before writing any code.