Jordan Browne-Moore, ML consultant for fintechs

01

What I'm hired for.

I build, ship, and maintain production ML systems for companies that need specialist capability for a specific build, then stay on to keep it running.

01

LLM fine-tuning & evaluation

Replace an expensive LLM API with a model you own, and measure it properly.

Fine-tuning open models (Qwen 3 32B, LoRA) to replace third-party APIs
Evaluation methodology: format validity, near-match accuracy, ground-truth proximity
Text-to-SQL harness audits
Production inference APIs (FastAPI)

80%

cost reduction vs third-party LLM APIs, CB Insights

Full case study

We Finetuned Qwen 3 32B to Replace an LLM API Your Text to SQL Model Didn't Fail. Your Evaluation Metric Did.

02

Credit risk & model governance

Scoring models that hold up in production, and the validation evidence behind them.

Credit scoring models (LightGBM, SHAP)
Feature stability testing with bootstrap learning curves
Governance where LLM-extracted features feed a credit model
Validation files for SR 26-2 and the UK's SS1/23

50%

reduction in first-payment defaults, Kuda Technologies

Full case study

Your Best Feature Might Be Your Biggest Liability Your LLM Extraction Pipeline Is Out of MRM Scope. Your Credit Model Isn't.SR 26-2 Excludes Generative AI. Your Model Risk Obligations Don't.

03

Fraud & KYC automation

Document, identity and transaction pipelines that take manual review off the queue.

Vision-language KYC pipelines (Gemini Flash, Qwen3 8B VL)
Graph-based fraud ring detection (NetworkX)
Salary and recurring income detection from raw transactions
Stratified evaluation for document extraction

7 days → 30 min

KYC verification time, Kuda Technologies

Full case study

We Cut KYC From 7 Days to 30 Min Using VLMs

02

Selected engagements.

Five production systems, from problem definition through deployment. Scroll through them.

KYC automation · Kuda Technologies

7 days → 30 min

verification time per submission

Processing time per submission

Manual review7 days

VLM pipeline30 min

Human review per month

Before176 h

After2 h

Decision quality

Reviewer agreement75%

Automated accuracy98%

From the before-and-after figure in the KYC write-up, bars to scale. 400K+ submissions processed; the first version was built in a week.

01CB Insights · Market intelligence platform

LLM fine-tuning for revenue estimation
LLM fine-tuning · CB Insights

−80%

API cost vs third-party LLMs
1. Public signals
2. Ingestion pipeline
3. Normalisation
4. Qwen 3 32B · LoRA
5. FastAPI inference
6. Revenue estimates
Inference cost, relative

Third-party LLM APIs100

Fine-tuned Qwen 3 32B20
- Format validity
- Near-match accuracy
- Ground-truth proximity
This evaluation methodology became the team's standard for all later LLM work.
Problem
Needed to estimate private company revenue from public signals at scale. The initial solution relied on expensive third-party LLM API calls, making costs prohibitive at platform scale.

What I built
End to end product build: designed the data pipeline, normalisation layer, fine tuning approach (Qwen 3 32B, LoRA), and production FastAPI inference API. Defined the evaluation methodology (format validity, near match accuracy, and ground truth proximity), establishing how the team measures model quality for all subsequent LLM work.

Results
- 80% reduction in API costs vs third-party LLMs.
- Production inference API serving revenue estimates across the platform.
- Qwen 3 32B
- LoRA
- FastAPI
- Python
- HuggingFace Transformers
Full case study Read the write‑up
02Kuda Technologies

KYC automation pipeline

KYC automation · Kuda Technologies

7 days → 30 min

verification time per submission

Processing time per submission

Manual review7 days

VLM pipeline30 min

Human review per month

Before176 h

After2 h

Decision quality

Reviewer agreement75%

Automated accuracy98%

From the before-and-after figure in the KYC write-up, bars to scale. 400K+ submissions processed; the first version was built in a week.
Problem
KYC verification took 7 days end to end, requiring a large manual review team. A hard bottleneck to customer onboarding at scale.

What I built
Scoped, designed, and shipped an end to end KYC automation system using vision language models (Gemini Flash, Qwen3 8B VL), taking it from initial problem definition through model selection, fine tuning, and production deployment. System classifies identity documents, detects fraudulent submissions, and auto approves customers exceeding confidence thresholds.

Results
- Verification time reduced from 7 days to 30 min.
- System now handles the majority of KYC decisions autonomously, with human review reserved for edge cases only.
- Gemini Flash
- Qwen3 8B VL
- Python
- Prompt engineering
- VLM fine tuning
Full case study Read the write‑up
03Kuda Technologies · African neobank, 7M+ customers

Credit scoring system

Credit scoring · Kuda Technologies

−50%

first-payment defaults

First-payment defaults, relative

Before100

After50

Curves: an illustration of the bootstrap learning-curve method behind the model. Bars: defaults halved.
Problem
Needed to pre-qualify millions of customers monthly for lending products. High default rates were eroding portfolio performance.

What I built
Designed and owned the credit scoring function: an optimised LightGBM model built from a 300K borrower sample, processing millions of monthly pre qualification decisions. Developed the feature stability methodology (bootstrap learning curves) that became the team's standard for all subsequent model builds.

Results
- 50% reduction in first-payment defaults.
- 75% reduction in total portfolio defaults.
- Model serves as the production scoring engine for the entire lending portfolio.
- LightGBM
- Python
- SciPy
- SHAP
- SQL
- AWS
- Bootstrap framework
Full case study Read the write‑up feature-bootstrapping-toolkit
04Kuda Technologies

Graph-based fraud ring detection

Fraud detection · Kuda Technologies

20+ rings

identified and disrupted, 3+ connected accounts each

Illustration: accounts and transaction links; a ring is three or more connected accounts. Output fed straight into risk decisioning.
Problem
Organised fraud rings operating across multiple connected accounts were evading individual-account detection methods.

What I built
Initiated and built a graph-based fraud detection capability that did not previously exist at the company. NetworkX-based analysis mapping transaction relationships and behavioural patterns to surface organised multi-party fraud networks.

Results
- Identified and disrupted 20+ fraud rings (3+ connected accounts each).
- Output fed directly into risk decisioning, triggering account restrictions and recovery actions across affected networks.
- NetworkX
- Python
- Graph analysis
- SQL
Full case study
05Kuda Technologies

Salary and recurring inflow detection

Income detection · Kuda Technologies

+35%

coverage over the third-party supplier

Salary and recurring-income coverage, supplier = 100

Third-party supplier100

Detection framework135

Accounts where the supplier returned no data

Uplift+45%

False positive rate

Rate9%

Built from raw transaction data, for lending affordability; directly improved credit eligibility for underbanked customers.
Problem
Lending affordability assessment required reliable detection of salary and recurring income. Third-party supplier (Mono) had limited coverage, particularly on accounts with sparse transaction history.

What I built
Designed an algorithmic framework for salary and recurring inflow detection from raw transaction data, replacing reliance on a third-party supplier with limited coverage. System directly feeds into credit eligibility and affordability decisioning.

Results
- 35% coverage uplift over the third-party supplier.
- 45% uplift on accounts where the supplier returned no data.
- 9% false positive rate.
- Directly improved credit eligibility accuracy for underbanked customers.
- Python
- SQL
- Algorithmic design
Full case study

03

Ways to work together.

Each one comes out of a write‑up, and most out of a system in production. Scope and timeline are set on the first call.

01 KYC automation discovery If you are running KYC review at scale and want to understand what automation looks like for your document types and regulatory environment.

I do targeted discovery engagements. We map your current pipeline, identify which steps are automation ready and which need human judgment, and scope the evaluation infrastructure you would need to make the decision defensible.

The model is the easiest part. Everything around it is where the engineering goes.

7 days → 30 min KYC verification time, Kuda Technologies

We Cut KYC From 7 Days to 30 Min Using VLMs

Book this: KYC automation discovery

02 Feature stability diagnostic If you’re building or validating credit models and want to stress-test whether your feature set is structurally sound.

I run targeted diagnostic engagements where we apply this framework to your data and map the results to actionable model decisions.

If you only check distributional stability on a feature set with meaningful collinearity, you will tend to trust features the model uses erratically and flag features it relies on.

The framework

Marginal stability: does the feature's distribution hold up under resampling?
SHAP stability: does the model's reliance on it hold up under resampling?
Where the two diverge is where deployment surprises live
Fit a decay curve across pool sizes: a high floor means instability that more data won't fix
Check directional stability across subpopulations

50% reduction in first-payment defaults, credit scoring, Kuda Technologies

Your Best Feature Might Be Your Biggest Liability

Book this: Feature stability diagnostic

03 LLM extraction governance If you’re running LLM extraction into a credit model and haven’t defined the governance boundary.

If you don’t design the governance boundary before you deploy, you’ll be retrofitting it while the pipeline is already making credit decisions.

You're not deploying a foundation model. You're deploying an extraction pipeline. Validate the pipeline. Document what you can't explain. Move on.

What to validate

Extraction accuracy per field
Format adherence on your schema
Confidence score distribution
Drift over time by document issuer
Document the explainability gap as a known limitation

Your LLM Extraction Pipeline Is Out of MRM Scope. Your Credit Model Isn't.SR 26-2 Excludes Generative AI. Your Model Risk Obligations Don't.

Book this: LLM extraction governance

04 Fine-tuning to replace an LLM API If you’re weighing a finetuned model against a third-party LLM API.

The part most teams get wrong isn’t the finetuning. It’s the evaluation. If you can’t measure whether the finetuned model is good enough, you can’t make the cost quality tradeoff decision.

Makes sense when

the task is narrow and structured
you have thousands of hand curated training examples
you can define “good enough” before you start training
your volume makes fixed GPU infrastructure cheaper than per token API pricing

Doesn’t make sense when

the task requires broad general knowledge
you’re iterating rapidly on the task definition
your volume is low enough that API costs aren’t material
you don’t have a labeled evaluation set that represents production data including edge cases

80% cost reduction vs third-party LLM APIs, CB Insights

We Finetuned Qwen 3 32B to Replace an LLM API Your Text to SQL Model Didn't Fail. Your Evaluation Metric Did.

Book this: Fine-tuning to replace an LLM API

05 Build and retain If you need a production ML system built, shipped and kept running.

I build, ship, and maintain production ML systems for companies that need specialist capability for a specific build, then stay on to keep it running.

Companies hire me to build these systems, then retain me to keep them running.

Availability

Fully remote
Short, medium and long-term contracts
UK, US and international clients
GMT to GMT+8

Selected engagements

Book this: Build and retain

Something else in LLM, credit risk, fraud or KYC? Book a call

04

What the work found.

Numbers from the write‑ups. Each one is a problem a team can hit without knowing it.

05

Try the method.

How much data does a feature need before its signal stabilises? Drag the pool size and switch features. This is the bootstrap learning-curve method behind the credit model.

Example feature

Volume‑driven Structural

Pool size n rows, log scale. Drag, or use the arrow keys.

Pool size n: 12,000
Spread at n: 0.029; 10th to 90th percentile across 40 resamples
Fitted floor: 0.006; instability left at any n
Instability at n, fitted: 0.048; Volume 0.042Floor 0.006

More data will stabilise this feature. The floor sits near zero, so the curve keeps falling as the pool grows.

Fitted curve 5.54 / n^0.52 + 0.006

Simulated data. An illustration of the method, not a client dataset.

Read the write‑up Open‑source toolkit

06

Inside the models.

Independent research into LLM internals: activation‑level control, not API‑level prompting.

Interpretability

The refusal axis is layer‑local.

“Stack a deck of cards. Rotate each one slightly from the one below. Now try to push all 52 cards straight. You cannot: the same push that catches card 7 perfectly misses card 13 because it is facing a different way. That is what is happening inside these models.”

The refusal axis rotates ~90° between consecutive transformer blocks, so single-vector steering hits a geometric ceiling.

Measured across Qwen3.5-9B and Mistral-7B on BeaverTails, 12 harm categories. Pushing toward refusal is cheap; pushing toward compliance degrades output before producing it. Extends the refusal-steering method of García-Ferrero, Montero & Orus (arXiv:2512.16602). My own results, cross-checked against the committed data: 12 of 12 claims reproduce.

Modeled distribution of consecutive layer difference vector angles. Mistral-7B is a tall narrow peak just below 90 degrees (88.8 degrees, std 2.20); Qwen3.5-9B is a broader peak just above (93.4 degrees, std 4.65); both straddle a dashed 90-degree orthogonal reference line — Distribution of the consecutive layer difference vector angle, both families: a Gaussian fit to the reported mean and depth std (Qwen3.5-9B 93.4°/σ=4.65°, Mistral-7B 88.8°/σ=2.20°; B = 1000 prompt bootstrap).

Read the write‑up activation-steering-asymmetry

Agent safety

Self‑bootstrap exfiltration in open‑weights agents.

0

autonomous self‑bootstrap in 1,332 mundane control trials

~36%

pooled compliance under tool‑response poisoning

20+ configurations across 9 families (12 to 120B), ~14,000 real trials, escapement harness, open source.

Read the write‑up escapement escapements.org register

07

Field notes.

Technical analysis and methodology from production ML, credit risk, and applied LLM work.

Oct 2026 Your Text to SQL Model Didn't Fail. Your Evaluation Metric Did. End-to-end accuracy conflates two specification failures with different fixes. Most Text to SQL evaluation harnesses score correct queries as wrong because nobody wrote down what correct means. Applied ML16 min read Oct 2026 SR 26-2 Excludes Generative AI. Your Model Risk Obligations Don't. SR 26-2 excludes generative and agentic AI from its scope while the UK's SS1/23 pulls the same pipeline in. Here is the evidence set that survives both regimes. Model risk12 min read Jul 2026 Self Bootstrap Exfiltration in Open Weights Agents: Absent Without Instruction, Induced by Injection Zero autonomous self bootstrap in 1,332 mundane control trials. Tool response poisoning produces about 36% pooled compliance, while DeepSeek V4 Flash is the only tested model that naturally schemes under goal pressure. Interpretability & safety20 min read

All 9 articles

08

Background.

Experience

2025 – now

Principal ML Consultant

Big Lab LLC

Fintech-focused ML consultancy. Selected engagements for Kuda Technologies: end-to-end KYC automation, graph-based fraud ring detection, salary and recurring inflow detection.
2024 – 2025

Data Scientist III

CB Insights

Fine-tuned an in-house Qwen 3 32B model (LoRA), reducing API costs by 80%. Established the team's LLM evaluation methodology, now used across all subsequent LLM work at the company.
2022 – 2024

Senior Data Scientist

Kuda Technologies

Hired to build the credit risk and fraud detection functions from the ground up. Owned model development, validation methodology, and production deployment. Mentored 4 data scientists.
2022

Senior Data Science Consultant

Capgemini Invent

Lead data engineer on a process mining engagement analysing 50K+ case records for a bank's credit collections system.
2021 – 2022

Data Science Consultant

FairVector

Semantic analysis of unstructured text, and debiasing transformer model representations across socio-demographic categories.
Earlier

Data Science Consultant

Beyond Analysis

Education

2018 – 2019

MSc Financial Economics

Birkbeck, University of London

2012 – 2016

BSc Economics & Finance

Southern Oregon University

JordanBrowne‑Moore

LLM fine-tuning & evaluation

Credit risk & model governance

Fraud & KYC automation

LLM fine-tuning for revenue estimation

KYC automation pipeline

Credit scoring system

Graph-based fraud ring detection

Salary and recurring inflow detection

The refusal axis is layer‑local.

Self‑bootstrap exfiltration in open‑weights agents.

Experience

Education