🛠️ AI Systems Builder & Quantitative Researcher · University of Oxford (Laidlaw Scholar) · N1 AI Scholar
🔬 Core Focus: AI Agents & Execution Environments · Quantitative Trading & Event Engines · LLM Training Dynamics · Econometric Data Pipelines

Python · PyTorch (Apple Silicon / MPS) · Transformer Mechanics · Controlled Empirical Study · Preregistered Recovery

• Research Problem: Does the temporal order of context lengths presented during training leave a persistent path-dependent deficit on an LLM's final capability, or is the apparent difference an artifact of recency bias and confounded evaluation?
• What I Built & Key Findings:
  • Engineered an inspectable 4.8M-parameter character-level Transformer optimized for Apple Silicon (MPS backend) with modular data pipelines, diagnostic probes, and evaluation harnesses.
  • Executed a three-iteration controlled empirical study: evolved from an initial single-seed pilot to a 5-paired-seed design, culminating in a formal preregistered recovery study holding token budgets (4,096 tokens/step), paired seed initializations, data manifests, and AdamW optimizer trajectories invariant.
  • Uncovered that an initial severe descending deficit ($+0.7570$ BPC gap at $T=256$) completely reversed to $-0.0485$ BPC after a common $T=256$ recovery phase—ruling out strong persistent path dependence and demonstrating that final performance is dominated by recent context exposure.
• 🔗 Repository | 📄 Concise Technical Report (PDF) | 📋 Frozen Preregistration

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

Python · Interactive Brokers (IBKR) · SEC EDGAR API · Streamlit · Probabilistic Valuation · Ongoing / WIP

• Research Problem: Corporate events (earnings releases, guidance updates, Investor Days) shift fundamental cash flows faster than equity markets price them, but naive sentiment approaches suffer from severe look-ahead bias and ungrounded hallucinations.
• What I Built & Key Edge:
  • Implemented an end-to-end quantitative trading engine calculating the Fundamental Revision ($FR$) vs Price Reaction ($PR$) Gap to capture structural underreactions across 10-minute to multi-day horizons.
  • Zero Look-Ahead Bias Ingestion: Pre-event snapshot validation strictly locking consensus expectations prior to event execution; live SEC EDGAR REST API reader with sha256 content hashing and primary citation tracking.
  • Quantified Probabilistic Valuation: Structured thesis engine computing probabilistic Bull/Base/Bear scenarios, Expected Value ($EV$), return spreads, and explicit thesis break conditions.
  • Automated Execution & Risk Limits: Interactive Brokers (IBKR) paper order routing protected by strict portfolio guardrails (12.5% single-stock cap, 50% portfolio cap, spread <1.5%) and human-in-the-loop Streamlit UI.
• 🔗 Repository | 📄 Architecture Specification

Python · AST Import Parsers · Invariant Baselines · Agentic Tool Execution · Safe Migration Pipeline

• Engineering Problem: AI coding agents frequently propose hallucinated structural changes, break module imports, and generate unsubstantiated novelty claims without verifiable environment feedback.
• What I Built:
  • Developed an enterprise-grade AI Agent Skill and deterministic verification environment for autonomous repository audits and high-conversion landing page restructuring.
  • 6-Domain Invariant Verification Baseline (scripts/invariant_checker.py): Programmatically captures and validates AST module imports, relative Markdown links, package entry points, and CI workflows.
  • AST Safe Migration Pipeline (scripts/safe_migrate.py): Performs dry-run simulations, atomic git mv operations, and instant automated git reset rollback on test or invariant failure.
  • 3-Layer External Novelty Audit: Orchestrates GitHub Search API, OpenAlex/arXiv API, and web search to output deterministic 6-part proof tuples (Claim -> Comparable -> Similarity -> Difference -> Evidence -> Confidence).
• 🔗 Repository | 📦 Skill Specification

• ⚡ gemini-live-multimodal-tutor: Real-time multimodal conversational tutor built with the Gemini Live API for voice/visual problem solving (Google Cloud Hackathon 2026).
• 🗺️ oxfordshire-population-map: Interactive geospatial portal mapping census demographic distributions and social enterprise density across UK postcodes (OX1–OX49) (Live Map).
• 📝 CrystalNotes: AI transcript intelligence pipeline transforming unstructured audio into deeply hierarchical, structured study notes.
• 🚀 tech-launch-promoter-skill: AI Agent Skill for generating developer launch threads, Show HN submissions, and 15-second screen demo scripts.
• 🌌 Datawhale_AI4S: Numba JIT-accelerated Kesten stochastic dynamics simulator & active learning candidate phase transition scanner.

• AI Agents & Verification Environments: Deterministic agent execution harnesses, AST-based import/dependency validation, 6-domain invariant checking, tool call schema verification, atomic migration pipelines.
• Quantitative Trading & Financial Engineering: Event-driven mispricing signals ($FR - PR$), SEC EDGAR live ingestion, abnormal return modeling, probabilistic scenario valuation ($EV$), Interactive Brokers (IBKR) API integration, risk management guardrails.
• LLM Training Dynamics & Mechanics: Context-length curricula, paired-seed experimental controls, preregistration protocols, BPC validation surfaces, Apple Silicon MPS profiling.
• Quantitative Econometrics & Data Pipelines: Multi-source REST/Bulk ETL pipelines, missing data theory (MCAR/MAR/MNAR), entity resolution (postcode-blocked Jaccard n-gram matching).
• Languages & Frameworks: Python (PyTorch, Numba, NumPy, SciPy, Pandas, Streamlit), JavaScript / TypeScript, HTML/CSS, SQL.

• 🎓 Affiliations: University of Oxford (Laidlaw Scholar) · N1 AI Scholar
• 🧠 nanoGPT Research: angelazu-builder/nanoGPT
• 📊 OSEP Research Portal: osep-quant-ai-social-impact
• 📈 Quant Trading Engine (Ongoing): event-driven-trading
• 🤖 Agent Environments: repo-organizer-skill
• 🌐 Interactive Maps: oxfordshire-population-map