01 / 05
The challenge
LendDesk was a two-year-old consumer lender processing about 400 loan applications per week through a mix of a legacy underwriting tool and a small team of analysts. Each application took 26 hours end-to-end. The founder's ambition was to double volume without doubling the team — which meant every step of the process needed to be automated, auditable, or eliminated.
There were three real bottlenecks: bank-feed analysis (tedious, human-driven), narrative summarization (analysts writing the same paragraphs over and over), and marginal-risk decisions (applications that weren't clearly approve or clearly reject, requiring senior-analyst attention).
02 / 05
Our approach
We built a Laravel origination platform as the system of record — every application, every document, every decision, every audit event. On top of it, we wired three focused AI capabilities: a fine-tuned bank-feed classifier (PyTorch, hosted in-house), a narrative-summary generator (OpenAI, prompted against a fixed schema with human review), and a marginal-risk explainer that surfaces the top three risk factors for any borderline application.
Critically, the AI never made the decision. Every automated recommendation was accompanied by an explainability panel; a human always signed the final approve/decline. This kept the founder's risk officer comfortable and, more importantly, kept us aligned with the regulator's guidance on model-assisted underwriting.
03 / 05
Key decisions
We hosted the bank-feed classifier ourselves. Sending a customer's full bank statement to a third-party LLM was not an option — both for regulatory reasons and because the founder had visible skin in customer trust. A fine-tuned open-source model on a small GPU box gave us 96% classification accuracy at a fraction of the runtime cost.
Every AI call was structured. No free-form prompts, no hallucinated fields. The narrative summary generator receives structured JSON and returns structured JSON, keyed against a schema. Guardrails catch drift; a nightly eval suite flags regressions.
We wired Chargebee for the subscription add-ons (premium repayment plans, insurance upsells). Chargebee's dunning ladder recovered 4.2% of at-risk revenue in the first month — pure margin.
04 / 05
What changed
Underwriting time dropped from 26 hours to 13.5 — a 48% reduction. Application volume doubled inside four months without adding headcount. The risk profile of approved loans is unchanged: default rate is holding within 2 basis points of the pre-automation baseline. PCI-DSS review completed with no blocking findings; SOC-2 Type 1 is scheduled for Q2.
05 / 05
Where we're headed
We're building a customer-facing lending assistant — a chat interface backed by the same AI infrastructure — that helps a prospective borrower understand what documents they need and why. Early experiments show it lifts application completion by 22%.
