Production ML, Not Notebook Demos
RAG systems, AI agents, and predictive models that ship to real users.
What I build
RAG Pipelines
Embeddings, retrieval with confidence scoring, citation surfacing. pgvector, LangChain, FastAPI. Served as REST API or embedded widget.
AI Agents with Real Tool Access
Multi-agent orchestration via CrewAI. MCP servers connecting agents to SQL databases, internal APIs, and documentation repos — not toy chains.
Predictive ML for Tabular Data
Classification, regression, and anomaly detection using CatBoost, XGBoost, LightGBM, and ensembles. Benchmarked against your baseline, not just cross-validated in a notebook.
OCR & Document Extraction
Structured JSON extraction from scanned PDFs. Azure Document Intelligence or in-house GLM-OCR depending on cost/accuracy tradeoffs.
Recent work
CompuCal ML Platform
Production ML features shipped inside CompuCal's regulated calibration-management platform. RAG over internal documentation, OCR pipelines for scanned certificates, predictive-maintenance models, and MCP servers giving AI agents real database access.
ReviveMap
3rd place at the Google Hackathon. A map-based platform that helps communities identify and revive neglected public spaces, with AI-assisted categorisation of user-submitted reports.
FinBERT + XGBoost Stock Prediction
Combines FinBERT sentiment extraction from financial news with gradient-boosted price-movement models. Benchmarked against a naive baseline with walk-forward validation rather than random splits.
NeuroMind System
Multi-agent orchestration built on CrewAI. Agents coordinate research, planning, and execution steps with tool access to external APIs rather than isolated prompt chains.
Stack
ML & AI
- • PyTorch
- • scikit-learn
- • CatBoost
- • XGBoost
- • LightGBM
- • pandas
- • NumPy
GenAI & NLP
- • LangChain
- • CrewAI
- • MCP
- • FinBERT
- • Transformers
- • OpenAI API
- • Gemini
Backend
- • Python
- • FastAPI
- • Node.js
- • TypeScript
- • Golang
Cloud & DevOps
- • Azure
- • AWS
- • Docker
- • GitHub Actions
- • Linux
Databases
- • PostgreSQL
- • pgvector
- • SQL Server
- • Redis
- • MongoDB
How I work
I work async-first and document as I go. Paid 2-hour scoping call before any engagement over 20 hours so we both know the scope is real. Fixed-price for well-defined deliverables, hourly for open-ended exploration. Based in Cork, Ireland — EU business hours with meaningful overlap for US east-coast mornings.