Work

Projects, with the numbers to back them

Each case study shows the problem, the approach, the measured result, the stack and the code. Sources for every number are on each project page.

All projects

• Content moderation classifier, with a fairness check

  Problem

  A single accuracy number can't say where to draw the line, how much lands on human reviewers, or whether a model is fair.

  Approach

  Trained a TF-IDF baseline and DistilBERT on 1.8 million public comments, set two threshold policies, simulated a review queue and checked false-positive rates by identity group.

  Result

  ROC-AUC 0.957; 36.6% of toxic comments auto-actioned at ≥ 90% precision; harmless comments wrongly flagged cut from 12.8% to 9.6%.

  • Python
  • scikit-learn
  • PyTorch
  • DistilBERT
  • Kaggle
  • Hugging Face
  • ONNX
  • transformers.js

  Case study: Content moderation classifier, with a fairness check · Demo · Code

• Dublin bikes + weather data pipeline

  Problem

  The Dublinbikes feed only shows right now; questions about weather and time of day need history that can be trusted.

  Approach

  A collector every 15 minutes, then a daily dbt + DuckDB run into an incremental star schema with 14 automatic data tests and Telegram alerts.

  Result

  Every station, every 15 minutes, modelled and tested into a live dashboard and public dataset, for €0.

  • Python
  • GitHub Actions
  • dbt
  • DuckDB
  • Parquet
  • Hugging Face Datasets
  • GBFS
  • Met Éireann

  Case study: Dublin bikes + weather data pipeline · Code

• Ask the EU Digital Services Act

  Problem

  People ask about the EU Digital Services Act in everyday words, while the law uses its own terms across 93 articles.

  Approach

  Split the official text into 482 chunks and compared keyword, semantic and hybrid search on 25 hand-written questions; the demo quotes the law only.

  Result

  The right article is in the top 3 for 84% of test questions (72% in the free browser version).

  • Python
  • BeautifulSoup
  • sentence-transformers
  • BM25
  • transformers.js
  • Kaggle
  • Hugging Face
  • GitHub Actions

  Case study: Ask the EU Digital Services Act · Demo · Code

• SkillBot: from job title to the skills it needs

  Problem

  Job ads and course catalogues live apart, so "what do I need to learn for this job?" is usually guesswork.

  Approach

  MSc research on an LLM plus a knowledge graph, and a free in-browser rebuild that matches a job title to similar public job ads.

  Result

  About 5 of 10 suggested skills appear in the real ad, in a leave-one-out test on 1,167 ads (48%).

  • Python
  • LLMs
  • Knowledge graphs
  • Prompt engineering
  • sentence-transformers
  • transformers.js
  • Hugging Face
  • GitHub Actions

  Case study: SkillBot: from job title to the skills it needs · Demo · Code

• Irish rents, 2008–2025: an open dataset, cleaned

  Problem

  Ireland's official rent table has 674,352 rows, codes repeated as labels and 70% empty cells.

  Approach

  A reproducible Python cleaning pipeline with a Kaggle notebook that rebuilds the data from the official source.

  Result

  201,912 rents for 437 places, 2008–2025, published on Kaggle and Hugging Face.

  • Python
  • pandas
  • Kaggle
  • Hugging Face Datasets
  • GitHub Actions
  • CSO open data

  Case study: Irish rents, 2008–2025: an open dataset, cleaned · Demo · Code

• Blog automation bot with Telegram approval

  Problem

  Cross-posting by hand to four sites is slow, but nothing should go public under my name without me seeing it first.

  Approach

  A Python bot on GitHub Actions that previews each post on Telegram and publishes only after one tap.

  Result

  Posts go to the blog, DEV and LinkedIn in about a minute of my time, on GitHub's free tier.

  • Python
  • GitHub Actions
  • Telegram Bot API
  • DEV (Forem) API
  • LinkedIn Posts API
  • Jekyll

  Case study: Blog automation bot with Telegram approval · Code

• Gmail auto-reply with Telegram approval

  Problem

  Routine emails take time, but an AI that sends email on its own can misunderstand or be manipulated.

  Approach

  Claude drafts a reply as a Gmail draft and sends it to Telegram with Send and Discard buttons.

  Result

  A routine reply becomes one tap, and the AI never sends anything by itself.

  • Python
  • Claude API
  • Gmail API
  • Telegram Bot API
  • Prompt engineering

  Case study: Gmail auto-reply with Telegram approval · Code

• Employment Permits 2025 dashboard

  Problem

  International students in Ireland rarely know which employers sponsor work permits, or when hiring peaks.

  Approach

  Cleaned the official 2025 permits dataset, grouped companies into sectors and built a single-page Chart.js dashboard.

  Result

  A searchable, serverless dashboard of permits by sector, month and employer.

  • Excel
  • Data cleaning
  • Chart.js
  • JavaScript
  • Tailwind CSS

  Case study: Employment Permits 2025 dashboard · Demo · Code