I take AI systems from mathematical modelling to production.

Senior AI Developer

Consultancy Outfit (CO-UK)

• Loophole Larry – RAG pipeline for a UK-law case solver: ingestion, indexing and retrieval optimization for queries that are whole case files hundreds of pages long. PII anonymization and redaction with Microsoft Presidio for UK GDPR.
• CossmoneyX / CurrencyGram – Designed and pitched to the CEO a 13-agent FCA compliance system: 11 parallel analyst agents, a coordinator and a reviewer producing a readiness report with strengths, gaps and improvements.
• Property Peak – Scoped nine AI agents across six user roles with role-based access and GDPR constraints; building them as LangGraph agents behind FastAPI with platform data via MCP.

Associate Team Lead – AI

Devsinc

• Voice AI – Led a real-time conversational video/voice agent with Speechmatics and Deepgram STT, RAG and LangGraph agents for multi-step dialogue.
• Latency – Migrated from a Pipecat telephony prototype to the ElevenLabs API, bringing end-to-end latency to a consistent 950 ms p95; replaced a custom web-search tool with OpenAI's built-in web search.
• Deployment – Deployed on AWS with production monitoring.
• Mentoring – Trained junior and intern data scientists on LLM development, code review and deployment.

Data Scientist

A1 Strategez

• Text-to-SQL – Fine-tuned Llama 3.1 8B so business users could query inventory data in plain English, raising accuracy on the client's dataset from 35% to 88% with a small, cheaply hosted model.
• ChippyGPT – Co-developed and deployed the backend for a ChatGPT-style assistant customised for a client's electronic-component data: proprietary prompts, their own database, and control over model selection and output format.
• Schema unification – Text-processing pipeline that unified heterogeneous inventory schemas and cut LLM token use by routing simple cases through classical NLP.

Math / Algorithms Expert

Zigron

Solar power generation & optimization – Terrasmart (US)

• Sole developer and architect of the end-to-end engine; reported directly to the Director.
• 3D model of site layout, sun position over discrete time steps and inter-row shading; optimized panel tilt angles for total daily energy output.
• Constrained optimization with MATLAB Global Optimization Toolbox solvers (pattern search, particle swarm, genetic algorithm), tuned for faster convergence.
• Cut runtime by vectorizing, parallelizing with the Parallel Computing Toolbox and porting hot paths to C++.

AI dentistry – automated analysis of dental X-rays

• Initial sole ML developer; demonstrated viability of AI for tooth numbering, diagnosis and bone-loss estimation.
• Annotated several hundred dental images; 90% accuracy on tooth numbering.
• R&D on DETR and on CLAHE / wavelet preprocessing for low-contrast X-rays.
• Deployed on a local server; presented capabilities and limitations to academics at University College Dublin.