From Naval Architecture to AI: Why I Changed Everything
This is not a technical post. It is an account of how I ended up studying Data Science and AI at the University of Liverpool after four years of Naval Arc...
Featured Projects
arXiv RAG System
Hit Rate 60→100%
MRR 0.51→0.82
async pipeline
End-to-end RAG for academic papers, built and optimised from scratch.
FastAPI
lifespan startup, fully async I/O, and a documented
28pp fine-tuning regression traced to training-data contamination.
FastAPI · ChromaDB · LLM (Qwen3 4B / Ollama) · Docker
FinScope — Multi-Agent Analyst
3-agent LangGraph
~15s parallel
hallucination check
Multi-agent RAG over SEC EDGAR & Companies House filings.
Retriever → Analyzer → Critic pipeline, parallel analysis via
asyncio.gather, and a Critic Agent with a conditional retry loop.
LangGraph · Groq (llama-3.3-70b) · ChromaDB · FastAPI · Langfuse
DefectVision
100% AUROC (bottle)
normal-only training
real-time webcam
Real-time manufacturing defect detection with PatchCore — trained on
normal images only, no labelled defects. FastAPI inference API plus a
Streamlit dashboard with live webcam streaming.
Anomalib · PatchCore · PyTorch · OpenVINO · FastAPI · Streamlit
TORCS RL Racing Agent
Recent Posts
LoRA Fine-Tuning Qwen3 4B for RAG: A 6.8-Hour Experiment That Failed Honestly
This post is a deep dive into the fine-tuning experiment from the arXiv RAG System. That post summarised it in a section - this one documents every detail...
arXiv RAG System: Async Refactoring and Bug Fixes
This is a follow-up to arXiv RAG System: Engineering an Academic Paper Q&A System from Scratch. The system was functionally complete after 7 days, but...
arXiv RAG System: Engineering an Academic Paper Q&A System from Scratch
arXiv RAG System: Engineering an Academic Paper Q&A System from Scratch
TORCS Corkscrew Challenge: A Journey Through Reinforcement Learning Failures and Breakthroughs
TORCS Corkscrew Challenge: A Journey Through Reinforcement Learning Failures and Breakthroughs