Darren
Wong

Live ML service — every prediction scored against the actual departure

An end-to-end delay-prediction service for Hong Kong International Airport, built on three public data feeds: 33 point-in-time features — rolling airline and airport delay statistics, as-of METAR weather joins, typhoon signals — feed XGBoost trained on a date-ordered split, reaching AUC 0.661 and a 16.6-minute MAE on a held-out 6,252-flight test set. A 30-minute GitHub Actions cron publishes predictions and a self-grading loop scores each one once the flight departs; scores export as static JSON, so the React + deck.gl dashboard is served entirely from static hosting.

Live dashboard →

A data-contamination study — does the agent forecast, or remember?

A no-lookahead, event-driven backtester — decisions at close t, fills at open t+1, 15 bps round-trip costs — replaying an LLM trading agent over fifteen US large-caps (2019–2026) against classical strategies. The agent turns out to be a high-turnover momentum chaser that trails simple equal-weight rebalancing risk-adjusted (1.04 vs 1.23 Sharpe) after a 3.5 pp/yr cost drag at 9.4× annual turnover. A controlled A/B — identical engine, prices and seed, real tickers versus anonymised labels — isolates data contamination: the real-name arm also cited sector vocabulary never supplied in the prompt. 107 offline tests and a committed prompt cache replay every number for $0.

Six deep-learning architectures — three detectors and three classifiers — trained and evaluated head-to-head on the Kaggle crop pest dataset, with an end-to-end pipeline covering YOLO-to-COCO conversion, early stopping, explainability visualisations, and a deliberately imbalanced split to test robustness. Completed on exchange at UNSW Sydney.

169 tests — 98 unit and 71 integration, the integration suite running against a real PostgreSQL container rather than an in-memory stand-in — over a schema managed entirely by Flyway migrations, with JWT authentication and role-based access across three user types. Started as a six-person team project and was substantially reworked afterwards. The domain is an ordinary one: catalogue, borrowing, reservations and fines.

AI Engineer Intern

CarryAI Company Limited, Hong Kong

• Ran the measured evaluation programme for the company’s vision-language safety monitor: labelled 650+ frames of real Hong Kong construction CCTV and showed a detect-then-describe pipeline — YOLO person crops fed to the VLM — lifts climbing-over-rail recall from 4/15 to 14/14 frames, while menu-of-states prompt design cuts false alarms from 181 to 4.
• Built a 232-photo IoU benchmark for a structural-steel inspection client that lifted rusted-bolt detection precision from 0.25 to 0.70.
• Rebuilt the company’s Prompt Helper into a client-testable product, starting from 133 mined production conversations — 80% of abandonments came before users answered a single question: an instant-draft interview, multi-frame rule building with negative-evidence frames, auth and CI across 43 commits, plus a standalone demo app with 58 tests.

Software Engineer Intern

Total Rehabilitation Management (TRM HK), Hong Kong

• Engineered a workflow analytics dashboard with data aggregation, multi-criteria filtering and interactive visualisation, letting stakeholders monitor, track and export K2 process metrics across both active and completed cases.
• Architected a document intelligence pipeline using TF-IDF vectorisation that reached 95% classification accuracy on unstructured document types — incident reports, medical certificates, receipts.
• Cut regulatory-report image payloads by up to 89% with lossless compression, keeping within third-party platform constraints, and hardened public endpoints with reCAPTCHA verification.

COMP & CPEG Mentor

HKUST Department of Computer Science and Engineering

• Advise first- and second-year computer science students on study planning, academic strategy and co-curricular involvement.

Peer Companion

HKUST Counselling and Wellness Centre

• Facilitate small-group activities supporting social skills and peer relationships, and organise university-wide wellness events.