🎓 Laidlaw Scholar & Researcher at the University of Oxford
🔬 Research Focus: AI for Science (AI4S) · Econophysics & Stochastic Dynamics · Active Learning & Agentic Systems · Quantitative Social Economy

Python · Numba JIT · Kesten Stochastic Dynamics · Active Learning (UCB+GP) · World Bank API Calibration

• Research Problem: Identifying non-continuous candidate phase transition thresholds in high-dimensional Kesten wealth dynamics ($W_{t+1} = A_t \cdot W_t + B_t$) without high-dimensional grid search bottlenecks.
• Key Innovations:
  • Implemented a Numba JIT-accelerated simulator ($N=100,000$, $T=500$) with Active Learning Agents (UCB + Gaussian Process Surrogate) to discover candidate phase transitions at $R^* = p/c \approx 0.15 \sim 0.20$.
  • Conducted Finite-Size Scaling ($N \in [5k, 100k]$) convergence proofs, Hill/Newman MLE tail fitting, and Bootstrap 95% Confidence Intervals.
  • Discovered two emergent phenomena: Absorbing Boundary Structural Collapse ($R^2 < 0.50$) and Keynesian MPC Power-Law Breakdown ($c(W) \sim W^{-\alpha}$).
  • Calibrated parameters with World Bank API macro data for US and China historical Gini coefficients.
• 🔗 Repository | 📄 Research Report (PDF/Docx) | 🌐 English Readme

Python · OpenAI Responses API · Missing Data Econometrics · Multi-API Data Pipeline · Folium / Leaflet

• Research Problem: Regulatory invisibility and missing data across micro-enterprises make evaluating social economy density and policy shocks challenging.
• What I Built:
  • Standardized 3,096 master entities across 7 REST/Bulk APIs (Companies House, Charity Commission, FCA, 360Giving, Contracts Finder, IMD) using postcode-blocked Jaccard n-gram matching ($\ge 85%$).
  • Evaluated OpenAI Responses API (web_search tool) vs Chat Completions (gpt-4o-mini), demonstrating a 96% vs 32% accuracy jump and zero-hallucination web extraction.
  • Modeled the 2013 CIO Policy Shock (+15,600% surge in CIOs, 61% drop in CLGs) and missing data theory (MCAR vs MAR/MNAR).
• 🔗 Repository | 🌐 Live Interactive Maps | 📄 Empirical Paper

JavaScript · Gemini Live API · Multimodal Audio & Vision · Google Cloud Hackathon 2026

• Problem: Traditional e-learning interfaces lack conversational adaptation and real-time visual problem-solving feedback.
• What I Built: Built for Google Cloud Hackathon 2026, an interactive multimodal AI cognitive companion leveraging the Gemini Live API for real-time voice and visual interaction during complex problem solving.
• 🔗 Repository

TypeScript · LLM Systems · NLP · Productivity Engineering

• Problem: Unstructured audio transcripts are time-consuming to review and lack clear hierarchical structure.
• What I Built: An AI-powered transcript processor that extracts core insights and formats them into deeply-layered, structured bullet notes.
• 🔗 Repository

Python · Folium / Leaflet.js · Geospatial Data Science · Spatial Modeling

• Problem: Communicating spatial variations in population density and social enterprise distribution across UK census postcodes (OX1–OX49).
• What I Built: Standalone interactive spatial mapping portal featuring choropleth layers, income quintile circle markers, and ward boundary tooltips.
• 🔗 Repository | 🌐 Live Map Portal

• AI for Science & Econophysics: Kesten Stochastic Processes, Candidate Phase Transition Scans, Active Learning (GP+UCB), Finite-Size Scaling, Hill MLE Estimators.
• Quantitative Research & Econometrics: Missing Data Theory (MCAR/MAR/MNAR), Structural Selection, Elasticity Modeling, Policy Shock Econometrics.
• AI & Agentic Systems: OpenAI Responses API (web_search tool), Gemini Live API, Structured JSON Outputs, Antigravity Agent Skills.
• Geospatial & Data Engineering: Multi-Source REST/Bulk ETL Pipelines, Postcode Blocking Fuzzy Matchers, Folium/Leaflet.js Spatial Analytics.
• Languages & Frameworks: Python (Numba, NumPy, SciPy, Pandas, Matplotlib), JavaScript/TypeScript, HTML/CSS, SQL.

• 🎓 Institution: University of Oxford
• 🌐 Interactive Maps: oxfordshire-population-map
• 📊 OSEP Research Portal: osep-quant-ai-social-impact
• 🌌 AI for Science Project: Datawhale_AI4S