MSc Data Science & AI graduate. I build and evaluate machine learning systems end to end: from data pipelines and NLP applications to reinforcement learning environments and model evaluation.

A task planner that calls its own tools, step by step

• Tech Stack: Python, Streamlit, Anthropic Claude (tool use), regex-based rule planner
• Highlights: two interchangeable planners (deterministic and Claude-powered), transparent reasoning trace, safe AST-based calculator, zero setup required to try it

A custom 2D rocket landing environment trained with reinforcement learning

• Tech Stack: Python, Gymnasium, Stable-Baselines3 (PPO), Streamlit, Plotly
• Highlights: physics built from scratch, hand-tuned control baseline for comparison, PPO agent trained from scratch in minutes, mission replay dashboard with live animation

Local document Q&A with page-level citations

• Tech Stack: Python, Streamlit, sentence-transformers, Hugging Face Transformers (extractive QA)
• Highlights: retrieval-augmented answers grounded in the source PDF, confidence scoring, chat-style interface, nothing leaves the machine

Real-time brand sentiment intelligence dashboard

• Tech Stack: Python, Streamlit, Hugging Face Transformers (RoBERTa), Plotly
• Highlights: live Google News ingestion, transformer sentiment scoring with TextBlob fallback, interactive dark dashboard, CSV and HTML report export

NLP application that analyzes university AI policies and generates evidence-based recommendations

• Tech Stack: Python, Flask, spaCy, scikit-learn, Plotly, SQLite
• Highlights: multi-format document processing (PDF, DOCX, TXT), 4-dimensional ethical framework analysis, recommendations grounded in academic literature, tested on policies from 9 leading universities

Two-part healthcare analytics project

• Analysis: ML study with feature engineering and model comparison, documented in a technical report
• Web App: Flask application with patient management, MongoDB/SQLite storage and a test suite

Designing reinforcement learning environments and the graders that score agents on them, with a focus on tasks that are genuinely hard, honest to grade, and resistant to shortcuts.

• Custom MuJoCo control - modified continuous-control robots where a controller tuned for the textbook model does not transfer, so the agent has to learn the new dynamics
• Hidden-environment, budget-limited tasks - the agent reaches the simulator only over a socket, one step at a time, under a hard interaction budget: sample-efficient learning rather than reading the answer
• Calibrated continuous scoring - deterministic graders that map measured return onto a 0-1 score with principled floor and ceiling anchors instead of arbitrary thresholds
• Reward-hacking resistance - degenerate-policy probes (all-zeros, constant, random) anchored at the floor, so nothing scores well without real skill
• Expert reference policies - SAC and PPO baselines trained on GPU with Stable-Baselines3, exported into the exact grader stack so calibration reflects what the grader really sees

• 🦾 Reinforcement learning: training and evaluating agents (SAC, PPO, Stable-Baselines3, Gymnasium/MuJoCo), designing custom environments and evaluation pipelines
• 🧭 Agentic AI: tool-calling planners and LLM-driven task execution
• 🧪 AI model training and evaluation for frontier AI systems
• 📈 Expanding my data science portfolio with end-to-end applications
• 💼 Open to freelance data science & AI projects

• 💼 LinkedIn
• 📧 info@jaai.co.uk

💡 "Transforming complex data into actionable insights through ethical AI and evidence-based solutions."