All work
MSc dissertation
AtlasExtract
A web application that turns public web pages into structured, validated datasets by generating, executing, validating, and repairing reusable extraction recipes.
Dissertation buildLLM systemsData infrastructureProvenance
Case file
signalmethodevidence
The problem
Research constantly needs structured data from public pages that were never designed to provide it — and one-off scrapers break silently when pages change.
What I built
- Three-pass LLM pipeline: vision (Groq Llama 4 Scout) for page understanding, two Cerebras passes for container selection and row extraction.
- Deterministic scraper executing strict, versioned extraction recipes.
- Validation with bounded automatic repair, then a relaxed-DOM retry with failure context.
- User-assisted fallback: click any card on the live page and the system derives the container.
- Provenance store: recipe versions, page snapshots, and validation reports.
My role
Sole builder — MSc Information Technology individual development project, University of Glasgow.
Methods
Pipeline designPrompt engineeringSchema validationBrowser automation
Why it matters
Deterministic execution wrapped around probabilistic components — the same evidence-first discipline as the rest of the work.
What it demonstrates
- LLM system design
- Deterministic execution around probabilistic parts
- Research ethics in tooling
Outcome
URL and a plain-language description in, validated dataset out — with every recipe version, page snapshot, and validation report kept for provenance.