William Catt · Data Scientist · Melbourne, Australia

From ambiguous question to working ML product.

I take unclear problems through data preparation, modelling, evaluation and deployment. My work spans NLP, retrieval-augmented generation, LLM evaluation, predictive modelling and simulation — built with Python, SQL, Azure and AWS.

Explore selected work Download résumé

Open to junior and associate Data Scientist and Applied ML opportunities in Melbourne or remote.

01 · Selected work

Three systems, with the evidence left visible.

Each case study covers the problem, data, architecture, evaluation, results and what did not work — plus source or a live demo where it can be shared.

01 · NLP / PRIVACY Evaluated prototype

Legal Text Anonymiser

A RoBERTa named-entity recognition pipeline that finds sensitive facts in legal judgments, then separates irreversible redaction from reversible pseudonymisation. The case study treats re-identification as a risk to measure, not a promise a model can erase.

0.851 partial-match F1 on the held-out annotation set

86.4% overall recall at the selected inference configuration

View Showcase Go to writeup

Pseudonymisation example showing tokenised legal text sent externally while the identity mapping stays in a local vault

Choose view

02 · RAG / LLM EVALUATION Evaluated prototype

QuoteGuard

A guardrailed retrieval-augmented generation system for insurance policy questions. It combines hybrid retrieval, reranking, quoted evidence and layered prompt-injection defences, then evaluates retrieval and safety separately.

0.803 recall@5 on 61 answerable evaluation questions

72.5 → 100% safe-rate on the 40-prompt tuned adversarial suite

View Showcase Go to writeup

QuoteGuard showcase answering a flood-cover question with a page citation, refusing personal advice, and displaying the guardrail trace

Choose view

03 · AGENTS / SQL Live local demo

Touchline

An agentic football analyst that translates questions into SQL, queries structured international match data and explains the answer with its evidence. Model comparison makes the quality, latency and cost trade-offs explicit.

49,291 completed matches in the demo database

30 questions across the model evaluation set

Read the case study

Touchline chart comparing answer accuracy for a frontier model and two local language models

02 · Process

The same delivery loop, adapted to the risk.

The output may be a classifier, a retrieval system or an analytics agent. The discipline is consistent: define what good means, test it, and keep limitations visible.

Frame

Turn an ambiguous request into an intended user, decision and testable outcome.

Prepare

Profile, clean and structure data while recording provenance and leakage risks.

Model

Start with credible baselines, then add complexity only when evaluation supports it.

Evaluate

Measure task quality, failure cases and operational constraints — not one headline score.

Deliver

Package the work as a usable product, API or reproducible analysis and monitor the trade-offs.

03 · Capabilities

Skills shown in context.

No rating bars: these capabilities link to the work, evaluation choices and constraints that demonstrate them.

NLP · Privacy

Information extraction under asymmetric risk

RoBERTa NER, annotation analysis, entity-level errors and privacy-aware output modes for sensitive legal text.

Anonymiser evaluation →

RAG · LLM evaluation

Retrieval and guardrails measured separately

Hybrid search, reranking, grounded responses, RAGAS/MLflow evaluation work and adversarial test design across portable and Azure implementations.

QuoteGuard results →

Modelling · Delivery

Data products with operating constraints

Predictive modelling and simulation in Python, analytical querying in SQL, and product delivery using Azure and AWS services.

Touchline comparison →Project archive →

04 · About

Technical curiosity, grounded in delivery.

I’m a Melbourne-based data scientist with a Master of Data Science from Monash University and industry experience in insurance. I enjoy the part of a project where an imprecise question becomes a measurable system — especially when NLP, retrieval or simulation is involved.

I care about reproducible evaluation, honest limitations and interfaces that let someone inspect the result rather than simply trust it.

Download résumé View all projects

05 · Contact

Have a data problem worth making concrete?

I’m available for conversations about junior and associate Data Scientist or Applied ML roles. The quickest way to reach me is email.

Email William LinkedIn ↗