← Selected Work
Regulated Financial InstitutionFinancial ServicesEmbed2024 — present

Credit intelligence in a regulated environment.

GenAI applied to lending decisions where every output must be explainable, auditable, and defensible to a regulator.

Credit decisioning
Duration
9 months, ongoing
Team
Embedded pod of 5
Core stack
Python · React · K8s
Models
GBM scoring + LLM rationale
Data
6M applications
Compliance
SOC 2-aligned · full audit
The challenge

Lending decisions were slow, manual, and inconsistent — but in a regulated environment a black-box model was a non-starter. Every decision had to be explainable and audit-ready.

Our approach

We embedded with the risk and data teams and built explainable decisioning with a human in the loop and an immutable audit trail — speed without surrendering control or compliance. Every model carries a model card and a documented validation process.

What we built

A decisioning workbench: applicant view, model rationale in plain language (SHAP-backed), override controls with reason codes, and an immutable audit log for every decision.

Tech stack
Application
TypeScriptReactFastAPI
AI / data
PythonXGBoostSHAPMLflowPostgreSQL
Models
Gradient-boosted scoringLLM rationale generationHuman-in-the-loop
Infra
Kubernetes (on-prem)Triton InferencePII tokenisation
What we delivered
  • An explainable decisioning workbench
  • SHAP-backed rationale for every decision
  • Override workflow with reason codes
  • Immutable audit log & model cards
  • A model-validation & monitoring pipeline
Timeline
4 weeks
Discovery with risk & compliance
8 weeks
Explainable model & audit framework
9 months
Embedded build & rollout
Ongoing
Monitoring & model governance
Data at work
6M
Historical applications modelled
100%
Decisions with an audit trail
0
Black-box outputs in production
Outcomes
Mins
Decision time, down from days
100%
Audit coverage
0
Unexplained decisions
Next case →Marketing automation at scale.

Have something
that needs to ship?

Tell us the problem. We'll tell you honestly what makes sense — even if that's not us.

Start a conversation