AI Engineering
Production LLM systems, RAG pipelines, multi-agent platforms. Each shipped behind real traffic with stateless services, instrumented endpoints, and evaluation harnesses.
LLM API Integration Microservice
An MCP server that gives any AI coding assistant instant access to live model discovery, pricing comparison, and ready-to-use connection snippets across seven providers — Anthropic, Google, OpenAI, DeepSeek, Z.ai, MiniMax, and Inception Labs. Built as a stateless HTTP microservice behind nginx, it plugs into Claude Code, Gemini CLI, and OpenAI Codex with a single URL — no API keys, no SDK installs, no configuration drift. One endpoint turns "which model should I use?" into an answered question with working code in five languages.
RAG Decision MCP
A Docker-first MCP server that gives AI agents structured, explainable guidance for building RAG systems — which architecture fits a workload, how to chunk each modality, and what retrieval stack, vector store, and evaluation setup to use at a given scale. A curated, versioned knowledge base drives deterministic, read-only recommendations with typed contracts and Mermaid architecture visualizations — no scraping, no runtime LLM calls. Hosted publicly: one URL, no API key, works identically across Claude Code, Gemini CLI, and OpenAI Codex.
The Architect
Systems Architect is a browser-based studio for system design interviews. It pairs a drag-and-drop canvas — built around a 7-layer architecture model with concrete tool choices at every layer — with live, realtime interview sessions: interviewers invite candidates by link, watch them design in real time, and pass canvas control back and forth. The AI layer (Gemini, OpenAI, or Claude — your key, your provider) generates calibrated challenges, scores designs against a six-dimension rubric, publishes hints, and walks candidates through a model solution step-by-step. Solo practice and live interviews share the same canvas, so what you rehearse alone is exactly what you run with a candidate.
ML Interactive Labs
Interactive machine-learning education platform spanning nineteen subject areas — from the original Policy Playground (Q-Learning, SARSA, bandits, multi-agent RL) through classic ML, search, deep learning, and Bayesian models. Every lab is a live, analytic, client-side simulation with real-time mathematical analysis, wrapped in a full-screen Cinematic Stage UI with a docked multi-provider AI tutor (Gemini, OpenAI, Anthropic, or DeepSeek — bring your own key). No backend, no servers.
Document Interaction System
RAG-powered Q&A system using Docling for multi-format document extraction, LangGraph agents with Google Gemini 2.5 Flash, and ChromaDB vector search. Dockerized with GPU acceleration, featuring markdown-aware chunking and streaming responses for intelligent document querying.
Neural Refresh
A React-based CS knowledge reinforcement app that helps developers combat skill atrophy through curated trivia questions across 10 domain categories, with optional AI-powered personalized learning plans via Gemini API.
Prompt Generator and Evaluator Agent
Web-based multi-provider prompt-engineering workbench. Generates platform-optimized prompts for Claude, ChatGPT, or Gemini and refines them with a dual-agent evaluator-revisor loop that prevents self-evaluation bias. Runs on Google, OpenAI, Anthropic, or DeepSeek engines with bring-your-own API keys held only in the browser session, selectable evaluator engines, and per-request thinking-level control.
Applied ML & Data Science
Computer vision, time-series, and tabular models for industry partners — from agri-food and manufacturing to financial datasets and Kaggle competition entries.
Tabular Document Segmentation Pipeline
AI-powered table extraction pipeline using ResNet deep learning models, corner detection algorithms, and OCR to automatically segment and digitize tables from historical documents at scale.
ML Harness
A reproducible local workflow for labelled tabular datasets that compares dependable scikit-learn baselines on identical cross-validation folds, records dataset hashes and environment metadata, and generates a fitted model with standalone training and prediction scripts ready to keep and extend in a repository.
Word Detection and Extraction Pipeline
Object detection pipeline using YOLOv7 for detecting and extracting words from historical document images with automated cropping, standardization, and batch processing capabilities.
Column Segmentation Pipeline with GUI Fallback
Advanced computer vision system for automated document line segmentation and column extraction using ML-powered instance segmentation, OpenCV, and interactive GUI tools for historical document digitization.
PDF Conversion Pipeline at Scale
Production ready PDF-to-TIFF converter with configurable dimensions, batch processing, and comprehensive error handling application with 90% memory optimisation.
Automated XGBoost Trainer for Kaggle
Automated Kaggle ML pipeline with intelligent preprocessing and XGBoost optimization for both Classification and Regression tasks with Streamlit user interface.
Regression and Classification Models
Jupyter Notebooks for 2025 Kaggle Playground competitions, covering regression and classification based tasks such as Podcast Listening Time, Calorie Expenditure, and Optimal Fertilizer Prediction.
Software Development
Tools and infrastructure I build to make the rest of the work tractable — full-stack TypeScript apps, C++ libraries, and Docker-deployed services.
Impact Flow
Production-ready full-stack task management system designed to measure and visualize work impact through revenue tracking, efficiency metrics, and analytics dashboards. Features recurring tasks, calendar/timeline views, and strategic OKR alignment. Deployed with Docker, secured with a defense-in-depth security architecture, and backed by PostgreSQL.
Orintu
A web-first interactive mathematics platform that teaches scenario-first: manipulate a real situation, commit an answer, then meet the formal theory and prove understanding through transfer rather than repetition. Lessons ramp through four rigor tiers — basics to research — and the whole curriculum is data over a small library of reusable interactive primitives, organised as a prerequisite DAG that recommends but never blocks. Optional AI grading of written proofs is client-side and bring-your-own-key (Anthropic, Google, OpenAI, or DeepSeek).
Traffic Control System
Red Means Go (RMG) - Traffic Simulation Framework. A C++ traffic flow simulation system designed to model vehicle behavior, traffic patterns, and road dynamics with a focus on extensibility, portability, and maintainability.
Head Pose Estimation System
Head Pose Estimation and Alignment System using Microsoft Kinect v2 and the Iterative Closest Point algorithm, developed for the Efficient Secure Access Control (ESAC) project at Lero.
Research Projects
Peer-reviewed work and ongoing investigations — Ph.D. thread on applied AI in agricultural and food systems, plus formal-methods and NLP collaborations.
Authored Research Papers
Peer-reviewed research publications spanning NLP, formal verification, and applied AI. Covers topics including Part-of-Speech tagging with HMMs, formal code verification with OpenJML, and probabilistic modelling for cybersecurity risk assessment.
Formal Verification with OpenJML
Formal verification case studies in Java using OpenJML, showcasing specification techniques and comparative verification approaches.
Hidden Markov Model for Part-of-Speech Tagging
A universal HMM-based POS tagger built to facilitate cross-linguistic machine translation through statistical language modelling.
AI Foundations
Reference implementations of foundational AI techniques — classical search, Bayesian networks, and from-scratch transformer notes I keep coming back to.
Best Path Optimization across State Space
A comprehensive Java implementation comparing fundamental search algorithms used in artificial intelligence and computer science applications. Developed and analyzed six core search algorithms across 9 test scenarios to evaluate performance characteristics including optimality, completeness, and computational efficiency for pathfinding problems.
Knowledge-Based Movement in State Space
D&C Sweeper is a Java-based implementation of a Minesweeper variant called "Daggers and Coins" that employs artificial intelligence algorithms to automatically solve game boards. The project implements various search strategies including Single Point Search (SPS), SAT Solver approaches, and random probing techniques.
Bayesian Alert System
A probabilistic threat detection system that models network attack risk using Bayesian inference. Maps vulnerability dependencies across enterprise infrastructure, enabling dynamic risk assessment as new evidence propagates through the belief network.