ML engineer. Singapore.
“the gunplay is tight but the story drags in the back half”
1,232 possible titles. No server.
I build machine-learning systems people can use, not just notebook experiments. The work often begins with a messy sensor stream, past papers or four million game reviews, then becomes something measured and shipped.
I am a third-year Computer Science and Design undergraduate at the Singapore University of Technology and Design, on exchange at the University of Waterloo for the autumn term. I previously spent four months at Bosch building time-series models for virtual sensing. I will join Infineon in December 2026 as a data science and automation intern.
The projects below are live or reproducible. The models on this site run in your browser, so you can try them without relying on a server or taking my word for it.
Data Scientist and Automation Intern · Singapore
Machine Learning R&D Intern · Singapore
Solo
Finds every past paper an exam question has appeared in, cites each appearance exactly, then marks a written answer one mark point at a time and explains why a correct-sounding answer still scored zero.
30,176mark points parsed from 315 papers
Industry project
A learning platform for the Dyslexia Association of Singapore that clusters learner assessment data into readable cohort profiles and generates differentiated literacy activities through a curriculum-grounded retrieval pipeline that refuses rather than invents.
Team of 8149 commits, ~16,700 lines
Sign in with the demo account shown on the login screen; public sign-up is closed.
Team of 5
Detecting machine-generated text when the test set is drawn from entirely different source corpora than the training set, which turned a model-selection exercise into a domain-shift one.
0.801Macro F1, up from 0.657; beats the shared task's fine-tuned RoBERTa at 0.7342
Solo
Two DistilBERT models over 4.2 million game reviews: one reads sentiment, the other names which of 1,232 games a review is about from the text alone.
60.7%top-1 across 1,232 titles, against a ~0.08% random baseline. Sentiment: 94.9% accuracy, 0.949 F1
UROP, self-initiated
An educational-content pipeline that turns notes into flashcards, a video and a PDF, implemented three separate ways behind one interface so the concurrency models could be compared directly.
3 runtimesThreadPoolExecutor, asyncio and Vertex AI ADK, over 253 Python and 45 front-end tests
Front end is deployed; backend hosting is still in progress, so generation does not run live yet.
Solo
ResNet50 transfer learning over mammogram and ultrasound images, retrained on a lesion-grouped split after an audit found leakage in the original evaluation.
97.09%leakage-free test accuracy on 962 images from 74 held-out lesions, against an 83.78% majority baseline; ROC-AUC 0.993
Solo
Predicting Indian crop yield from climate and agricultural-input data, where per-crop specialised models beat a single general one by a wide margin.
0.967adjusted R² over 42,306 records, measured on a fresh retrain against a held-out 20% split
Team
A recipe-discovery Android app with a short-video feed, allergen detection and per-recipe chat.
3rd best projectin the cohort, ISTD Information Systems & Programming, AY 2026
Solo
A discrete-time Markov chain over ARGOS satellite telemetry, validated on each shark's final observations against persistence, uniform and empirical-marginal baselines.
t=10: +7.6 pptop-5 hit rate over persistence (77.8% vs 70.2%) on a per-shark temporal holdout; tied at the one-day horizon
Team, 50.007 Machine Learning
Catching an evaluation artefact before it became the headline number: a prior group's F1 = 0.9973 did not survive an honest, sealed holdout.
0.653sealed holdout F1, positive class; the trustworthy number after 0.9973 was found to be an artefact
Solo
A proof-of-concept marketplace for giving unread physical books a second life: list a book, browse others, request an exchange, and rate the other side once it completes.
UROP, self-initiated
Generating a dataset of twisted bilayer photonic crystals via RCWA simulation, as the training data for a future surrogate model of circular dichroism spectra.
Work in progress. Stage 1 (data generation) is done; the forward surrogate and inverse design stages haven't started.
BEng Computer Science and Design
GPA 4.99/5.0. SUTD ASEAN Scholarship. Honours List AY 2024/2025.
Exchange semester, Computer Science (Faculty of Mathematics)
Canada-ASEAN SEED-2 Scholarship recipient (Government of Canada).
Available for an internship, May 2027 – Sep 2027.
The fastest way to reach me is owencliffton@gmail.com. I am also on LinkedIn and GitHub.