William Catt

Data science · Melbourne, Australia

I’m William.
A data scientist
in Melbourne.

I build tools for working with text and data: finding sensitive details in legal judgments, answering insurance policy questions, and exploring international football results.

Explore selected work Download résumé

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

01 · Selected work

A closer look at the work.

Start with a walkthrough, or read a case study for the technical decisions and results.

01 · NLP / PRIVACY Evaluated prototype

Legal Text Anonymiser

A tool I built to find and mask sensitive information in legal judgments. I tested what it misses and what the remaining details can reveal. The results show why human review is still needed.

127 unique judgments in the held-out evaluation
555 annotation rows across those judgments
View walkthrough Read case study
Pseudonymisation example showing tokenised legal text sent externally while the identity mapping stays in a local vault Interactive walkthrough
02 · RAG / LLM EVALUATION Evaluated prototype

QuoteGuard

An insurance-policy assistant that finds supporting clauses, cites its sources and explains when it cannot answer. The public companion lets you inspect the retrieval and safety tests separately from my Azure implementation.

0.803 recall@5 on 61 answerable evaluation questions
72.5 → 100% safe-rate on 40 adversarial prompts used to tune the checks
View walkthrough Read case study
QuoteGuard showcase answering a flood-cover question with a page citation, refusing personal advice, and displaying the guardrail trace Interactive walkthrough
03 · AGENTS / SQL Evaluated prototype

Touchline

Ask a question about international football and inspect the SQL, match records and charts behind the answer. I compared models on accuracy, response time and cost. The browser demo replays saved model plans and runs the SQL locally.

49,291 completed matches in the demo database
30 questions across the model evaluation set
Read the case study Try browser demo
Touchline chart comparing answer accuracy for a frontier model and two local language models Model comparison

02 · Process

How I work

I use the task and its failure cases to decide what to build and test.

01

Frame

Identify who needs the answer, what they will do with it and what a mistake would cost.

02

Prepare

Check where the data came from, what is missing and whether the test set is independent.

03

Model

Build a simple baseline so I can tell whether a more complex model helps.

04

Evaluate

Inspect errors alongside scores. For the anonymiser, that meant checking what private text remained.

05

Deliver

Make the result inspectable through an interface, source code and instructions for running it.

03 · Capabilities

Technical experience

Examples of the methods and tools I’ve used.

NLP · Privacy

Finding sensitive information

Fine-tuning RoBERTa for named-entity recognition, reviewing annotation quality and testing detection against full removal.

Anonymiser evaluation →
RAG · LLM evaluation

Search and language-model evaluation

Hybrid retrieval and adversarial testing in QuoteGuard; Azure OpenAI, Azure AI Search, RAGAS and MLflow in the professional implementation.

QuoteGuard results →

04 · About

A little about me

I have a Master of Data Science from Monash University and industry experience in insurance.

Download résumé View all projects

05 · Contact

Get in touch

If you’re hiring for a junior or associate data science or applied ML role, I’d like to hear from you. Questions about the projects are welcome too.

Email William LinkedIn ↗