Featured
Two MCP servers any AI coding assistant can call, and a realtime system-design studio with live interview sessions.
LLM API Integration Microservice
Discover models across ten providers, compare published prices and rate limits, and generate connection snippets in five languages — available to coding assistants through MCP and to applications through a shared catalogue endpoint.
RAG Decision MCP
Structured, explainable RAG architecture guidance for AI agents — deterministic and read-only, with no scraping and no runtime LLM calls. One public URL, no API key, identical across Claude Code, Gemini CLI, and Codex.
The Architect
One system-design canvas for real projects, interview practice and live candidate sessions — with AI review, cost estimates and a strict mock-interview mode. Interviewers can watch designs develop and pass canvas control back and forth.
Live Products
Deployed, publicly reachable, and running on infrastructure I provision and maintain myself.
ML Interactive Labs
Learn ML by running it — nineteen subject areas with live maths, adjustable simulations and an AI tutor. Now includes eleven RAG pipeline variants, a separate architecture viewer, and runnable Python exports of the configuration on screen.
Impact Flow
Task management that measures whether the work mattered — revenue tracking, efficiency metrics, and OKR alignment behind a defense-in-depth security architecture on PostgreSQL.
Neural Refresh
CS practice for working developers — on-demand quizzes across ten domains and over seventy subcategories, followed by learning plans for missed questions. Choose Google, OpenAI, Anthropic or DeepSeek; progress stays in the browser.
Prompt Workbench
Writes better prompts than you will by hand — a dual-agent evaluator-revisor loop that avoids self-evaluation bias, across four engines with keys that never leave the browser session.
Orintu
Mathematics taught scenario-first: manipulate a real situation, commit an answer, then meet the formal theory. Four rigor tiers from basics to research, organised as a prerequisite DAG that recommends but never blocks.
AI Infrastructure & Agent Tooling
Reproducible ML, governed coding agents, and tested model-catalogue rollouts across a self-hosted application fleet.
Pi Harness
A version-controlled operating environment for Pi — pinned tools and skills, permission hooks, shared rate-limit governance, local-model discovery and bounded context, with tested installation and rollback.
ML Harness
Makes a tabular baseline reproducible by construction — dependable scikit-learn models compared on identical CV folds, with dataset hashes and environment metadata recorded alongside standalone train and predict scripts.
LLM Fleet Updater
Keeps model catalogues consistent across a fleet of AI apps — one saved catalogue, per-app compatibility and cost policies, isolated validation, and sequential deployments with health checks and rollback.
Research
Peer-reviewed work and ongoing investigations, from formal verification and NLP through to mechanistic interpretability of deployment quantization.
Authored Research Papers
Peer-reviewed publications across NLP, formal verification, and applied AI — including HMM part-of-speech tagging for cross-linguistic machine translation, code verification with OpenJML, and Bayesian modelling of cybersecurity risk.
k-q Roundtrip
Makes real-world deployment quants measurable with interpretability tooling for the first time — a dequant round-trip loads GGUF k-quantized weights into a hookable PyTorch model.
Formal Verification with OpenJML
Proving Java code correct rather than testing it — specification techniques and comparative verification case studies using OpenJML and theorem provers.
In active development — links go live as soon as they're ready.
Briefbound
Formerly Founder OS: a desktop studio that turns an idea or existing repository into a reviewable architecture, component blueprints and build tickets. It hands deterministic build specs to coding agents, then reconciles their changes back onto the canvas. Local project files, your own AI keys or CLI subscriptions, and design, pricing and launch-planning tools connect the full build loop.
Sports Management System
A club and team platform for hurling, Gaelic football and camogie: role-based onboarding, squads, recurring fixtures, pitch bookings, team sheets and live scoring. Tactics Studio adds multi-phase formations, animated playback, 2.5D camera views and image/video exports. Web workflows are implemented; secure messaging is disabled pending independent review, with mobile and production readiness still in progress.
Earlier Work
The digitization pipeline below was production Ph.D.-era work. Everything else is listed for completeness.
Historical Document Digitization at Scale
An end-to-end computer vision pipeline that turns scanned historical archives into structured, queryable data — four production stages from bulk format conversion through table extraction and OCR.
- 01 PDF → TIFF conversion PDF Conversion Pipeline at Scale ↗
- 02 Column & line segmentation Column Segmentation Pipeline with GUI Fallback ↗
- 03 Word detection (YOLOv7) Word Detection and Extraction Pipeline ↗
- 04 Table extraction & OCR Tabular Document Segmentation Pipeline ↗
- Document Interaction System ↗ RAG-powered document Q&A with LangGraph agents and ChromaDB — superseded by the live RAG Decision MCP server.
- Automated XGBoost Trainer for Kaggle ↗ Automated Kaggle pipeline with intelligent preprocessing and XGBoost tuning for classification and regression, behind a Streamlit UI.
- Regression and Classification Models ↗ Notebooks for 2025 Kaggle Playground competitions — podcast listening time, calorie expenditure, and optimal fertilizer prediction.
- Head Pose Estimation System ↗ Kinect v2 head pose estimation and alignment via Iterative Closest Point, built for the ESAC secure access control project at Lero.
- Traffic Control System ↗ A C++ traffic flow simulation framework modelling vehicle behaviour and road dynamics, built for extensibility and portability.
- Bayesian Alert System ↗ Probabilistic network threat detection over a belief network of vulnerability dependencies — the basis of the cybersecurity publication.
- Hidden Markov Model for Part-of-Speech Tagging ↗ A universal HMM POS tagger for cross-linguistic machine translation — written up in the publications above.
- Knowledge-Based Movement in State Space ↗ D&C Sweeper — a Minesweeper variant solved automatically with single point search, SAT solving, and probabilistic probing.
- Best Path Optimization across State Space ↗ Six core search algorithms implemented and compared across nine scenarios for optimality, completeness, and efficiency.