Hi, I'm Wei Han.

Kuala Lumpur · MalaysiaAvailable for selected opportunities
Get in touch

About

I am drawn to the edge of what AI can do, rebuilding its best ideas from their papers so that one day I can add new ones of my own.

Backend Engineering

Distributed systems, from race conditions to safe rollouts and production debugging.

Model Architectures

Rebuilding Llama 3 and DeepSeek-V2 from their papers, and designing attention that costs constant time per token.

Computer Vision

A YOLOv8 detector that counts steel rebar from a single photo, at 96.82% counting accuracy.

LLM Applications

Retrieval-augmented question answering and Gemini-powered tools, with a human kept in the loop.

2022 – 2025Bachelor of Computer Science, Monash University MalaysiaGPA 3.41 / 4.00

Lab

A System One model, running in this tab.

A small model that runs entirely in your browser, on your own CPU through WebAssembly, with no server, API or GPU behind it. It's a 29M-parameter 4-bit quantised language model (15 MB), repurposed as a System One aimbot for DOOM.

Whenever an enemy comes into view, the model takes the aim and fires on your behalf.

WASD move←→ turn the crosshairSpace open doors, respawn, next levelP pauseF full screenLeft stick moveDrag on the game turn the crosshairUSE open doors, respawn, next levelTap the game show the full screen buttonTap outside pause
state sent to the model

—

Inspired by TypeSafe's Jev, we make a model perform quick actions without reasoning. Every four game tics, it reads the state of play as a single English sentence and answers two questions, whether to fire (a Noul) and which way to turn (a Choice).

Game: DOOM (shareware episode) by id Software

Writing

Written after the sleepless nights.

Projects

Highlights

    Experience

    Work

    Nov 2025 – PresentCurrentMelbourne, Australia · Remote

    LynkSphere

    AI & Software Developer

    • Co-building MailHQ, an AI outreach platform in Next.js and TypeScript.
    • Integrated Google Gemini to draft a personalised email per recipient, with human review and a server-side daily quota.
    • Secured the Supabase PostgreSQL schema with row-level security on every table.
    • Encrypted per-user Resend API keys with AES-256-GCM, with per-recipient open, click and reply tracking.
    • Built CSV contact import and a Recharts analytics dashboard, deployed on Vercel.
    Jun 2025 – Oct 2025CompletedKuala Lumpur, Malaysia

    Ant International

    Backend Engineer Intern · Blockchain Team

    • Diagnosed a race condition in a multi-region batch job's distributed lock from lock-database logs.
    • Rewrote the write path in Java and Spring Boot to insert first, closing the race and halving database round trips per statement.
    • Released through a 1% → 100% grey-scale rollout behind a runtime feature flag, with 0 production incidents.
    • Proposed a centralised scheduler microservice to remove duplicates by design, documented in a system-analysis report.
    • Validated a partner-bank payment integration, mocking bank responses from production traces when UAT was unreachable.
    • Documented in-house middleware systems as onboarding guides for new joiners.

    Skills

    Toolkit

    Languages

    PythonJavaTypeScriptJavaScriptKotlinSQLBashC++HTML/CSS

    Technologies

    Spring BootFastAPINode.jsNext.jsReactPostgreSQLMongoDBSupabaseRedisKafkaDockerKubernetesAWSGoogle Cloud

    AI & Machine Learning

    PyTorchTransformersLLMsRAGLangChainVector SearchYOLOv8Computer VisionHugging FaceScikit-learnNumPyPandas

    Tools

    GitGitHub ActionsJupyterVS CodeVercelRenderGemini APIClaude APIClaude CodeJUnitpytest

    Soft Skills

    Technical WritingRoot-Cause AnalysisStakeholder CommunicationCode ReviewAgile (Kanban)

    Leadership

    Leadership

    Bank Liaison

    Ant International2025

    Spoke for the team with partner banks during a payment integration, talking through the specification and the software decisions it raised.

    Onboarding Guide

    Ant International2025

    Led new joiners through in-house middleware systems, from RPC to dynamic configuration, with onboarding documents written for the purpose.

    Team Lead

    Monash University Malaysia2025

    Led a four-person final-year team in building an app that counts steel rebar from a single photograph, and took charge of training the model at its heart.

    Contact

    Have something worth building?

    I'm open to collaborations, research conversations and full-time opportunities in AI.

    Based in Kuala Lumpur · UTC+8