Matthew Paver · Portfolio Store
Open these first
One live product, one private system, one MVP. Each preview goes into the problem, how it is built, and where to inspect it.
Inference Brief
Live AI-news reader with paid subscriptions, bookmarks, and reading history.
Open preview
Happening
103+ venue pages turned into event data with checks you can reason about.
Open preview
QuickSupply
School cover requests, agency assignment, teacher offers, and live status.
Open preview
Catalogue mode
The store is organised like a working lab: live products first, then the data systems, automation, ML, analytics, and archive work that show how the ideas are built.
Live AI-news reader with accounts, paid subscriptions, bookmarks, and reading history. I collect, score, and summarise stories, then publish them as issues that subscribers actually use.
London event data is scattered across venue pages. This private pipeline gives each source explicit config, extraction rules, dedupe, checks, and scheduled runs.
Revision prep is the job here: upload notes, generate flashcards and quizzes, build study plans, then keep review moving through a spaced-repetition loop.
A signal layer for job-market research: listings are normalised, then salary, skill, remote-work, and volume changes become reports and alerts.
Supply cover needs coordination, not another spreadsheet. QuickSupply sketches the school, teacher, and agency flow around requests, availability, assignment, and live updates.
A property-operations prototype for the messy middle: resident requests, service-charge visibility, ticket audit trails, payment flows, and AI-assisted triage.
Built to show the full local analytics loop: raw marketing CSVs into DuckDB, clean tables, XGBoost models, data-quality checks, and a Streamlit dashboard.
Upload schedule data and get a sharper read on delivery risk: slippage, milestone pressure, forecast issues, and reporting outputs.
A compact image-generation project: curate the architecture dataset, train a conditional GAN, inspect generated outputs, and expose the model through a small Flask API.
A dashboard package designed for handoff: prepared data, PBIX files, screenshots, business commentary, and walkthrough material.
A small NLP notebook with a clear point: embeddings and cosine ranking are useful for retrieval, but similarity is not the same thing as truth.
Archive
Useful provenance kept out of the main shelf. These show earlier automation patterns without competing with the flagship builds.
Earlier automation work kept for context: article-to-newsletter rendering and image collection utilities that fed later product and dataset experiments.
Nothing matches that filter combination.
Clear the search or choose All to see every project.
Featured build
The hard part was not scraping one website. It was making 103+ different source shapes predictable enough to run, inspect, and trust.
Live deploy gate
This site only deploys when the spec-driven validator passes. Lighthouse CI gates accessibility and SEO. The panel reads the validator's last run.
Credentials