CV

Philbert Poku

MEng Computer Science Student, University of Southampton

University of Southampton — MEng Computer Science, on track for a First
Sept 2024 – June 2028

Relevant modules: Optimisation, Computer Architecture, Cloud Computing, Computer Vision, Foundations of Machine Learning

MacroPaths (OxHack '26) — Rust, Solana, TypeScript, Python — GitHub — October 2026

Co-designed, in a team of four, a Solana program in Rust that pays out on the path US inflation takes over 48 months: two periods starting and ending within 0.2 points settle at 83% and 29%.
Shipped the TypeScript keeper that commits 48 monthly CPI readings to a SHA-256 hash chain and replays them on Solana devnet without double-posting; fixed a high-severity bug that could lock every deposit.
Backtested the team's inflation forecast out of sample in Python with a rolling-origin test and an 11-origin walk-forward; it beat history on excess inflation but not on cooling, so the team stopped pricing cover on cooling.

Chessathon (Optiver x AI Chessathon) — Python, python-chess, NumPy, Numba — September 2026

Programmed a chess engine from scratch in Python on one CPU core and 2GB of RAM, with negamax alpha-beta search, and roughly doubled its search speed with Numba JIT compilation and incremental Zobrist hashing.
Automated A/B tests of every change across 20+ versions in an SPRT early-stopping arena; tuning one pruning threshold lifted the score against the same opponent from 49% to 76%, before 81 rated games against rival engines.

Dispatch (Cursor Cybersecurity Hackathon) — Python, FastAPI, Tauri, TypeScript — GitHub — August 2026

Built a security-remediation system: a FastAPI server has Cursor Agent draft sourced advisories from malware intelligence for an operator to approve, and a Tauri client for macOS and Windows checks each locally.
Simplified fixing to one vulnerability at a time in a single-panel interface with one "Run fix" action.
Isolated the AI agent from the verdict: a fix counts only when its captured final check classifies as clear, never on the agent's exit code, and a failed probe can never report a false all-clear.

Dissertation: INT8 Quantization of a LiDAR 3D Detector — PointPillars, KITTI — September 2026 – Present

Using an HPC cluster to investigate how INT8 quantization changes the accuracy of a LiDAR 3D object detector (PointPillars on KITTI), and which objects it loses, for teams deploying detectors on low-cost edge hardware.
Verified the field's headline claims against primary sources rather than summaries, establishing that the widely quoted accuracy collapse is a three-class mean driven mainly by cyclist and pedestrian failures, and catching an R11-vs-R40 mismatch in OpenPCDet's published baseline.
Found and fixed a selection-bias flaw before running any experiments: comparing only objects both precisions detect silently drops the objects INT8 loses, so the design scores every ground-truth object, counting misses as zero IoU.

Exchange Lab: Market Simulator and Strategy Research Platform — C++20 — August 2026 – Present

Engineered a price-time-priority exchange matching engine in C++20 with a lock-free input queue and zero allocation while trading, matching 99.95% of orders in under 2 microseconds and replaying bit for bit on every OS.
Developed a terminal app that scores any trading strategy against a do-nothing placebo in identical markets, using paired statistics, multiple-testing correction, held-out seeds and mutation-tested controls, so luck cannot pass as skill.
Extended it to real market data: it imports eight formats including LOBSTER, Databento and Nasdaq ITCH, and calibrates the simulation to resemble a file behind a known-answer check, so strategies are tested on realistic markets.
Quantified what reaction speed is worth to a market maker: 96% of a slow maker's loss remained with the informed trader removed, so it loses order flow to a faster rival rather than being picked off; failed controls were published.

Wildfire Watch: Smoke Detection and Response Swarm — Python, YOLOv8, NVIDIA NIM, Docker — May – June 2026

Created a wildfire smoke-detection pipeline in Python that runs YOLOv8 over hilltop camera time-lapses, turns each plume into one event and matches it to historical CAL FIRE incidents within 50 km by traceable rules.
Designed an MLflow-tracked YOLOv8 training pipeline on 2,192 smoke images, keeping the demo fires unseen.
Confined the Llama 3 model on NVIDIA NIM to writing summaries, never locating or confirming a fire, so a model error cannot trigger a response; a strong match sends a simulated agent swarm routing around the perimeter with A*.
Hardened it for unattended running with schema-validated events, a dead-letter queue, a circuit breaker and a recovery watchdog, shipped as a Docker image by a GitHub Actions CI/CD pipeline behind an 85% test-coverage gate.

Languages: Python, C/C++, TypeScript, Java, SQL, Rocq
Frameworks & Libraries: PyTorch, TensorFlow, NumPy, Pandas, OpenPCDet, OpenCV
Tools & Platforms: CUDA, MATLAB, Linux, Docker, Git