Everything I've learned shipping AI into lending.
Two to three posts a week. Opinionated, concrete, honest. No "ultimate guides," no stock metaphors, no "in today's rapidly evolving landscape."
Adverse action notices with LLMs: the explainability bar regulators actually enforce
How to draft an adverse action notice that survives ECOA Reg B, GDPR Article 22, and the plain-language fairness expectations of the FCA, MAS, APRA, OSFI, and RBI — with a working prompt and three jurisdictional renderings of the same decline.
Credit memo LLM bake-off: Claude vs GPT-4o vs Gemini on the same synthetic borrower file
A reproducible side-by-side run of Claude 3.5 Sonnet, GPT-4o, and Gemini 1.5 Pro on the same six-section credit memo prompt and the same synthetic borrower file — with scoring, costs, and an opinionated routing recommendation.
Early-warning signals for P2P lenders: the LLM repayment-message read that flags loans 30 days before they go bad
A weekly LLM review of borrower repayment messages and platform notes flags a meaningful share of the loans about to go 30+ DPD — earlier than the dashboard. With the prompt, a five-tier severity rubric, and a 30-thread benchmark.
AI credit scorecard vs ML model: which one actually fits a small-to-mid-sized lender?
The honest comparison between a rules-based AI-augmented scorecard and a trained ML default-prediction model — when to pick each, what they cost, and why the scorecard wins for most lenders under 10,000 active loans.
AI in loan collections: what actually works, what doesn't, and where compliance draws the line
A builder's view of where LLMs earn their seat in collections workflows — draft generation, segmentation, hardship conversations — and where they're a compliance liability.
Bank statement tampering detection with LLMs: the forensic prompt chain that catches surgical edits
A four-check LLM forensic chain — arithmetic, cadence, vendor fingerprint, PDF metadata — that catches roughly four out of five surgical bank-statement edits before disbursal, with honest false-positive numbers.
Portfolio allocation for retail P2P lenders: the weekly LLM review that catches concentration drift
A 15-minute weekly LLM review prompt that scores a retail P2P portfolio across five concentration axes and produces a watch-list of the loans most likely to default next quarter.
Underwriting thin-file borrowers with LLMs: the diligence prompt chain that turns alternative signal into a credit decision
A four-step LLM diligence chain that converts gig earnings, rent ledgers, and telco data into the same scorecard columns a thick-file applicant fills out — for borrowers in any market.
10 red flags AI catches in loan applications (that humans miss under time pressure)
Ten specific red-flag patterns that LLMs reliably surface from pasted application text — with the signal, the prompt logic that catches it, and the false-positive rate you should expect.
AI tools for collections: a teardown of what actually works
An opinionated review of AI-driven collections tools and workflows for small-to-mid lenders — what genuinely helps, what's vendor theatre, and where DIY prompts beat the SaaS.
Credit memo generation with LLMs: the prompt chain that survives a committee
A structured prompt chain that produces first-draft credit memos your credit committee will actually accept — with risk grade, conditions, deviations, and explicit gaps flagged.
P2P lending with AI: what actually works (and what's marketing)
What AI genuinely helps with when you're lending as an individual or running a P2P platform — borrower screening, portfolio allocation, collections — and the patterns that remain stubbornly human.
AI fraud detection for lenders: patterns, prompts, and the playbook
The four fraud patterns every small lender eventually meets — synthetic identity, document manipulation, ring fraud, and payout-stage scams — and where AI actually catches them vs. where it still misses.
How to underwrite loans with AI: a builder's guide (2026)
The full workflow for AI-assisted loan underwriting — from application intake to credit memo, with the prompts, scorecard logic, and failure modes that matter.