AI/ML engineer building optimisation, data, and full-stack software systems

GitHub • LinkedIn • Experience • Selected Work • Coursework

I'm studying for a BSc in Economics, Finance and Data Science at Imperial College London (2025–2028). I work at the intersection of machine learning, optimisation, data infrastructure, and production software engineering.

• President of Imperial College London's Economics, Finance & Data Science Society
• Google Student Ambassador (selected from over 300,000 applicants globally)
• Selected for Y Combinator's Paris Startup School 2026 (400 selected from 4,000+ applicants)
• Particularly interested in AI infrastructure, industrial ML, retrieval and knowledge systems, quantitative engineering, and reliable agentic software

Much of my recent company work is in private repositories. My public work spans Python services, PostgreSQL/Supabase platforms, retrieval and model systems, Next.js applications, optimisation, reinforcement learning, quantitative trading infrastructure, and systems tooling.

• A local-first AI software factory for sandboxed coding-agent campaigns, independent verification, and controlled promotion to commits and pull requests
• Multi-source institutional-memory platforms with versioned ingestion, provenance, hybrid retrieval, citations, and permission-aware access
• Multi-task NLP tagging and efficient inference systems, including transformer fine-tuning and ONNX deployment
• Reliable automation with explicit boundaries, human approval gates, strong tests, and recoverable execution

• Deep.Meta — Reinforcement Learning Engineer Intern (Jul–Sep 2026): Supporting development and evaluation of reinforcement learning and machine learning (ML) models for steel manufacturing optimization, with a focus on predictive systems, industrial time-series data, and digital-twin applications.
• Really Good Culture — Full-Stack Engineer Intern (Jul–Oct 2026): Improving ETL pipelines, supporting agentic orchestration products and dashboard development by building user-facing features, and contributing to multi-layer intelligence integrations across retail, brand, and macro/micro signals.
• EduNova — Software Engineer Intern (Jun–Aug 2026): Improved lesson history navigation by building filters in TypeScript/Next.js, synchronizing tables, summary widgets, and empty states. Automated contract generation with a LaTeX pipeline integrating PDF compilation, uploads, email delivery, progress updates, and testing

• EFDS Knowledge Platform — An end-to-end institutional-memory and AI platform for Imperial's EFDS Society: multi-source ingestion, PostgreSQL/Supabase, schema migrations, provenance, hybrid retrieval, embeddings, row-level security, a FastAPI agent service, streamed cited responses, and a Next.js/TypeScript application. Related repos: agent · site.
• Startup Knowledge Base — A multi-source knowledge system ingesting Slack, Linear, GitHub, Attio, files, and structured records into versioned PostgreSQL documents and retrieval units, with embeddings, hybrid search, provenance, an HTTP API, and MCP access.
• codex-autoapprover — A security-conscious Rust launcher for structured Codex CLI permission hooks, with child-local arming, exact compatibility gating, fail-closed behaviour, protocol tests, threat modelling, and isolated verification.
• Model-to-Market Trading Bot — A dry-run-first MT5 FX/crypto system with typed configuration, market-data collection, deterministic features, offline backtesting, pre-trade risk controls, SQLite storage, reconciliation, analytics, and guarded live execution.
• Mini-GPT with Rotary Embeddings — Built and pretrained a compact GPT on a Wikipedia corpus using rotary positional embeddings, improving benchmark performance by 2.3× over the standard positional-embedding baseline.
• Rocket State Estimation — Modelled a 19-state Kalman-filter estimator in MATLAB/Simulink, fusing LiDAR, GPS, RTK, IMU, barometer, and magnetometer data for real-time flight-state estimation.
• MetaTrader 5 Algorithmic Trading Platform — A Model to Market 2026 Hackathon finalist: an end-to-end Python platform integrating live FX and cryptocurrency data, signal generation, backtesting, risk validation, and guarded live execution. Its volatility-managed momentum strategy uses ATR-based exits, spread and slippage guards, portfolio exposure limits, drawdown controls, SQLite-backed evaluation, and a fully auditable execution pipeline.
• Retail Intelligence Dashboard — A retail analytics platform that transforms 130 consumer transcripts and product metadata into interactive, evidence-backed insights, with deterministic signal extraction and positioning-gap analysis across consumer voice, products, pricing, and brand context.

• Mini-GPT with Rotary Embeddings — Developed and pretrained a compact GPT inspired by minGPT on a Wikipedia corpus, using rotary positional embeddings to improve benchmark performance by 2.3× over the basic positional-embedding baseline.
• Rocket State Estimation — Modelled a 19-state Extended Kalman Filter in MATLAB/Simulink, fusing seven sensor inputs for real-time flight-state estimation and validating it against a physics reference model.
• Cherokee–English Seq2Seq NMT — Designed a neural translation system with a bidirectional LSTM encoder and unidirectional decoder, achieving a BLEU score of 11.77 on an extremely low-resource Bible corpus.
• MuJoCo Physics Simulations — Applied REINFORCE policy gradients with variance reduction to continuous-control tasks including hopping, running, and inverted-pendulum stabilisation.
• Warfarin Dosage Bandit Algorithms — Implemented and compared pharmacogenetic, clinical, and fixed-dose strategies using LinUCB, epsilon-greedy, Thompson Sampling, and other multi-armed bandit algorithms.

• Flareify — ETH Oxford 2026 — Developed a decentralised derivatives platform on Flare Coston2 supporting leveraged gas futures and stablecoin depeg protection, using Solidity contracts, a Next.js frontend, and Python/Node.js oracle infrastructure. Solana engineers later invited the team to continue pursuing the project through Superteam Ireland.
• Speech-Therapy.ai — HackEurope 2026 — Built an agentic, voice-first speech and language therapy application for people with stroke-induced aphasia, delivering accent-aware, personalised exercises through LLM orchestration, speech recognition, voice generation, and voice cloning.

These repositories contain implementations, experiments, reports, and LaTeX notes from Imperial and the Stanford AI Professional Program.