winston.liang
Toronto, Canada

I train models.I ship software.I fill rooms.

Reinforcement learning research at NUS. Data models at Shopify. A 400-person AI conference at UofT.

Speaking on stage at EigenAI, UTMIST's flagship AI conferenceEigenAIThe conference I ran400+ tickets · sponsors incl. Stripe

Third year at UofT, Computer Science and Data Science.

01 — the pathWhereI've been.

Software Engineering Intern

now
ShopifyToronto, ONSept 2026 — Present

Building the data models and internal tools Shopify's sales organization runs on, helping reps prioritize the merchants most likely to buy.

  • Improved the average precision of the merchant-ranking ML model by 90% across 800k training records, adding signals of merchant interest in Shopify's funding products.
  • Shipped a new field through two production data models, giving sales reporting a call, email, and review breakdown across 320K+ tasks and unblocking a dashboard migration off a legacy data source.

AI Research Intern

now
National University of Singapore · Multi-Agent Robotic Motion LaboratorySingapore · RemoteMay 2026 — Present

Researching how classical motion planning and reinforcement learning combine so agents can navigate shared space without colliding.

  • Trained a residual RL policy that layers real-time corrections on top of a multi-agent planner — 99.8% success rate, autonomous deadlock resolution, and 27% lower agent makespan.
  • Built a model that solved a benchmark map every standard grid-based planner failed on (0% solve rate across 64 test cases), delivering 100% success at 2.2x the speed and 10x fewer collisions than the industry-standard collision-avoidance method.
  • Diagnosed three distinct reward-hacking exploits in a multi-agent PPO pipeline — episode-prolonging reward farming, goal re-entry cycling, and a non-potential-based shaping loophole — then redesigned the reward function while preserving training stability.

Software Engineering Intern

Hong Kong Digital Policy OfficeCyberport, Hong Kong SARJul 2025 — Aug 2025

Automated vulnerability intelligence for the HKSAR government and prototyped on-premise LLM inference for sensitive internal data.

  • Cut vulnerability intelligence reporting from 2+ hours to under 1 minute (99.9% faster) with a Python scraping and alerting pipeline, distributing high-risk alerts to 70+ government bureaus and departments.
  • Built a local inference prototype with Ollama and Open WebUI reaching ~10 tokens/sec and 5 concurrent users on Apple M3 silicon, letting staff search internal documents without sending anything to external clouds.
  • Evaluated quantisation, distillation, and retrieval pipelines to run 3B–32B parameter models on constrained hardware for a 10+ user internal pilot.

Early Identification Program

CitiRemote, USMar 2026 — Apr 2026

Selected as 1 of 163 students across North America.

02 — the workThingsI built.

Drag, swipe, or use the arrows

01
Dec 2025 — Present

Scoliosis AI

Medical imaging, end to end

Full-stack AI web app that detects scoliosis from spinal X-rays with a YOLOv8 model trained on ~1,500 labelled medical images, at 96% precision.

Next.jsPyTorchYOLOv8FastAPIAWSPrismaPostgreSQL
02
Jun 2026

Drift

AI microlearning platform

Full-stack platform that turns any topic into a personalized learning track with LLM-generated syllabi, lessons, and feedback-driven progression.

Next.jsSpring BootKafkaMongoDBDockerAWS EC2Anthropic API
03
Apr 2026

Order Matching Engine

Multi-process exchange simulation

High-performance multi-process order matching system in C, using bi-directional UNIX pipes and I/O multiplexing for concurrent IPC between a virtual exchange and independent trader processes.

CUNIX IPCLinuxSystems
04
Apr 2026

Blackjack-RL

Play against a learned policy

Five RL algorithms behind a real-time training dashboard with configurable hyperparameters — play blackjack against the models and sharpen your probability intuition.

FastAPIReactGymnasiumVercel
05
Sept 2025 — May 2026

SceneClarity

UTMIST · AV perception reliability

Modular ML evaluation pipeline that scores how reliably autonomous-vehicle perception holds up in adverse weather. Transfer learning with ResNet-18/50 on WeatherNet (18,039 images) reached 89.48% accuracy and 0.897 macro-F1.

PyTorchTensorFlowJiraGoogle Colab
06
Apr 2026

UofT Course LLM

RAG academic advising

RAG advising system over verified course and policy data — text-embedding-3-small for semantic embeddings, ChromaDB for vector search, GPT-4o-mini for grounded generation. A backtracking constraint solver returns ranked, conflict-free 3–5 course timetables in ~6–16ms median.

Next.jsChromaDBFastAPISQLite
07
Nov 2025

TripBuddy

Clean architecture in Java

Java desktop trip planner built on SOLID principles and Clean Architecture. Authored 4,000+ lines of production Java across a 6-person team, leading REST API integration and the accommodation features.

JavaSwingREST APIs
08
2023

Virtual Tour

Hoi Ping Chamber of Commerce Secondary School

Panoramic virtual tour of the school campus built with panolens.js, stitched from Insta360 captures. Full-stack build under the school's IT Panel Head.

panolens.jsJavaScriptHTML/CSS
03 — try itPoke atthe work.

Three of these projects, running here instead of described.

Agents routing around you

Why multi-agent path-finding is hard

Draw walls on either grid. Both planners re-plan instantly.

Independent A*

Every agent plans alone

Conflicts now / total
0 / 4
Makespan
23
Unrouted
0

Prioritized space-time A*

Plans against a reservation table

Conflicts now / total
0 / 0
Makespan
25
Unrouted
0

Both are classical baselines, not my research policy. Prioritized planning is collision-free by construction, but it is incomplete — squeeze the map hard enough and watch Unrouted climb, because an agent that plans later can find every route already reserved. Getting past that limit is the problem I work on at NUS.

Play the blackjack agent

Dealer
??
You
??
Hands 0W 0L 0Push 0

Scoliosis detection

Spinal X-ray before inference
The same X-ray with YOLOv8 vertebra detections overlaid
rawdetections
YOLOv8 trained on ~1,500 labelled images. Each box is a vertebra with its confidence; the long box is the scoliosis classification. Frames taken from the project's own demo.
04 — toolsWhatI use.
PythonJavaCTypeScriptJavaScriptSQLBashAssembly (MIPS)PyTorchJAXTensorFlowKerasPPO / RLGymnasiumYOLOv8scikit-learnNumPyPandas
Spring BootFastAPIKafkaDockerAWS (EC2 · S3)PostgreSQLMongoDBGitHub ActionsCaddyLinuxReactNext.jsTailwind CSSPrismaVercelClerkGitJira
06 / honors

Scholarships & awards

10 scholarships and 16 competition awards, from university back through secondary school.

Scholarships

  • University of Toronto International Merit AwardAwarded for academic merit on admission.
    University
  • Dato Tong, Luis Charitable Foundation Limited ScholarshipSelected as 1 of 4 Hong Kong Scholars.
    University
  • Dean's List ScholarUniversity of Toronto — 4.0/4.0 academic year distinction.
    University
  • The Regents Participation AwardVictoria College — for a significant contribution to university student life.
    University
  • Kwan Nang Chong Scholarship — Student of the YearHoi Ping Chamber of Commerce Secondary School — role model student of the cohort.
    High School
  • Ho Kien Bong ScholarshipOverall 1st in cohort.
    High School
  • Wong Chung On ScholarshipOverall 1st in cohort.
    High School
  • Hoi Ping Chamber of Commerce Hong Kong ScholarshipAwarded for outstanding academic performance.
    High School
  • World-renowned University Admission ScholarshipAwarded for admission to a top-tier global university.
    High School
  • First in Class — 15+ subject scholarships over 6 yearsRanked 1st in Computer Studies, English Language, Liberal Studies, History, Economics, and Geography.
    High School

Awards

07 / involvement

Leadership & community

Student ML research, a conference, event programming for the UofT Hong Kong community, and a long run of building things for my high school.

VP Logisticsnow

UTMIST

Running logistics for UTMIST, including EigenAI — the team's flagship AI conference. Sold 400+ tickets and brought on sponsors including Stripe.

  • Sold 400+ tickets across the two-day programme.
  • Secured sponsors including Stripe.
Speaking on stage at EigenAI, UTMIST's flagship AI conference
Sept 2026 — Present

Machine Learning Developer

UTMIST

Built a modular ML pipeline for SceneClarity scoring autonomous-vehicle perception reliability. Ran EDA on EDD100k and WeatherNet, trained ResNet-18/50 in PyTorch, and evaluated with torchmetrics.

Oct 2025 — May 2026

AI Open Source Developer

UTMIST

Contributed to a Keras + NVDA screen-reader add-on for accurate image labelling, aimed at a friendlier browsing experience for visually impaired users. Focused on dataset sourcing, performance testing, and data cleaning.

Nov 2024 — Sep 2025

Director of Events

UT Hong Kong Students Association

Led a team of 6 planning events that bring people on campus together, including orientation day with 140+ participants and 40+ coordinated helpers.

May 2025 — May 2026

Event Executive

UT Hong Kong Students Association

Organized events reaching 100+ total participants, promoting Hong Kong culture in collaboration with other campus clubs.

Oct 2024 — May 2025

Student Developer

Hoi Ping Chamber of Commerce Secondary School

Built the school's panolens.js virtual tour, a visitor-registration system on Google Apps Script, and a point-of-sale system in C for the HKDSE public examination.

Sep 2021 — Aug 2023

Captain, Speech & Debating Club

Hoi Ping Chamber of Commerce Secondary School

Captain in 2022–23. Coordinated the team to 1st Runner-up in the Hong Kong Secondary School Debating Competition, and earned Gold for Debating at Yale and UCL in the World Scholar's Cup.

Sep 2021 — Jul 2023

Chairperson, Multi-Media Production Team

Hoi Ping Chamber of Commerce Secondary School

Nominated Outstanding Chairperson. Led ~60 students documenting school activities and coordinating production work.

Sep 2018 — Jul 2023

Chairperson, IT & STEM Club

Hoi Ping Chamber of Commerce Secondary School

Chairperson in 2022–23. Ran workshops and published computing explainers on the club's social channels.

2018 — 2023
08 / certifications

Certifications

Machine Learning SpecializationDeepLearning.AI · Aug 2025
Azure AI FundamentalsMicrosoft · Jan 2021
Introduction to Cloud ComputingIBM
Introduction to Software EngineeringIBM
Introduction to CybersecurityCisco · Jan 2021
05 — endSayhello.

Happy to talk about research, systems work, or a role you're hiring for.