Who we build for · Underwriting
J Labs builds the discipline behind AI-generated underwriting output: the system-prompting strategy, the reference library every prompt change is checked against, and the evaluation loop that keeps summaries consistent across every deal, not only the supervised ones.
Summary
J Labs builds AI systems for underwriting platforms and lenders that already generate summaries, scorecards or recommendations with a model and now need that output to hold up every day, across every deal, without a person checking each one. The work has three parts. The system-prompting strategy sets the structure, constraints and house style that anchor every summary the platform ships. The prompt-data library is a curated bank of reference items, organised by underwriting dimension, that every prompt change is checked against before it lands in production. The evaluation loop runs that check automatically, so a change that improves one dimension cannot quietly degrade another. For a US AI-native underwriting platform operating at scale, J Labs helped shape the prompting strategy and build a library of 84 reference items across 21 underwriting dimensions, and the system was production-ready two weeks after kickoff. The result is output lenders read and trust at a glance, with an audit trail behind every summary. Every engagement is fixed scope, fixed timeline and fixed price, from a four-day Spec at €5,000 to multi-phase builds above €200,000.
What we automate
The structure, constraints and tone that every summary follows, written down and versioned, so output reads the same whichever deal, analyst or model version produced it.
A curated bank of reference items organised by underwriting dimension. Every prompt change is checked against it before release, so the platform has evidence, not opinion, that the change helped.
Automated checks that run the library against a candidate prompt or model and report the dimensions that moved. A regression on one dimension blocks the release instead of reaching a lender.
Scorecard and deal summaries generated where the analyst already works, from the data already in the deal, with the reasoning available on demand rather than hidden.
Which prompt version, which references and which inputs produced each summary, recorded so a credit committee, an auditor or a regulator can be answered without a scramble.
Fit
This is for you if
AI output already reaches customers or a credit committee. The question is no longer whether to use a model but how to keep its output consistent and defensible at scale. That is the system J Labs builds.
Prompt changes ship on instinct. Someone edits a prompt, the summaries look better on three deals, and nobody knows what happened on the other thousand. A reference library and an evaluation loop end that.
You can name the dimensions that matter. Debt service, collateral, concentration, management: if your underwriting has a shape, the library can be built around it. Twenty-one dimensions were in scope for the US platform.
This is not for you if
You are still deciding whether to use AI at all. The evaluation discipline is for output that already exists. If the question is where AI fits in your underwriting, the honest first step is a 20 or 45-minute call, not a build.
You want a model trained from scratch. J Labs builds the prompting, reference and evaluation layer on top of the model APIs you already use. Training a foundation model is a different project with a different kind of team.
Nobody can sign off what a good summary looks like. The reference library needs an underwriter who can say this one is right and that one is not. Without that owner the evaluation has nothing to measure against.
Systems
The ones you already run. We don't insist on rip-and-replace: the automation layer goes on top of the systems in place, and if a system is genuinely past saving we say so on the discovery call and show our working.
Engagement options
| Engagement | Price | Timeline | What you get |
|---|---|---|---|
| The Spec | From €5,000 | Four days | A written brief you own outright: the dimensions in scope, what a good summary looks like, the acceptance criteria, an architecture sketch and a cost estimate. Vendor-neutral: take it to us or to anyone. |
| Prompt system and library build | Fixed price, quoted after discovery | Typically two to eight weeks | The prompting strategy, the reference library and the evaluation loop in production against your platform, with weekly check-ins, open staging, documentation written as we go and a post-launch support window. |
| Underwriting AI programme | Multi-phase, to €200,000+ | Multi-month, phased | Every product line and dimension on the same discipline: summary generation inside the platform, audit trail, and the evaluation gate wired into your release process, phase by phase. |
None of these is an off-the-shelf product. There is no J Labs software to install: every system is specified in the brief, built for your stack, and owned by you. Every engagement is fixed scope, fixed timeline and fixed price, with acceptance criteria agreed before work starts. See how we work and the Spec.
What we've shipped
AI-native underwriting · USA
84 reference items in the prompt library, 21 underwriting dimensions in scope, production-ready two weeks after kickoff.
Live availability
Live slots from our team calendar, weekdays, European working hours. A 20-minute intro or a 45-minute working session, your choice, on Google Meet. You get a one-page summary within 24 hours.
How much time do you need?
Step 1 · Pick a date
Loading availability…
Frequently asked
Or start with our brief templateWe reply within one business day.