Hello, I'm
Richael
Building AI that remembers
Year 11 student building GPU kernels, fine-tuned language models, and full-stack applications with real users, with a background in mathematics and modelling competitions.
My About
2026 – 2028 · Auckland
Year 11 Student
Graduating 2028 — IB Math AA HL and NCEA Level 1, planning AP Calculus and Computer Science exams next year — while building AI products on the side.
Competition
Mathematics & Modelling
Honorable Mention at HiMCM and the IMMC regional round, co-authoring modelling papers with a team under competition deadlines. AIME 10/15, AMC 12 Honor Roll (top 5%), and High Distinction in the Australian Mathematics Competition.
Ongoing
AI Research & Development
Using Claude, VS Code, and Modal to read frontier model papers, write and benchmark GPU kernels, fine-tune and serve language models, and ship web apps.
Volunteer
Worldwide Math Community
Volunteering in Mustang Math Community in Technology and Curriculum Development, contributing technical skills, creative thinking, and mathematical problem-solving experience.
Personal Details
- Name
- Richael
- Location
- Auckland, New Zealand
- Status
- AI & GPU Systems
- huangd6666@gmail.com
- GitHub
- github.com/dh8116
Currently Learning
Reading
Kimi K3 technical report
At Part 10 of a part-by-part read — through KDA and attention residuals, Stable LatentMoE, and into post-training. Part 11 is on building trustworthy rewards, which runs straight into the research below.
Building
Triton GPU kernels
One kernel a week, each benchmarked against PyTorch and published with its numbers. Best so far: fused cross-entropy at 15.90 ms against PyTorch's 24.03 ms at vocab 131,072 (1.51x), with 1.67x less peak memory.
Researching
RLVR verifier research
Leading a five-person paper on when a persistently wrong verifier actually harms GRPO training — holding error rate and asymmetry fixed and sweeping persistence. Targeting COLM / TMLR.
My Skills
My Projects
Soulor AI
Soulor is a companion app that remembers you. Memory and relationship stage reach generation itself, so a companion's voice shifts across five stages from stranger to soulmate — and the same engine runs simulation modes, for rehearsing a hard conversation before having it for real.
VNportal
VNportal is a vinyl and music platform built on real-time audio. Find a record, play what sits behind it, then mix it, cut it into a DAW, or play a rhythm game charted from its own detected beats — and generate single instruments, bar-locked to the tempo you set.
My Resume
View resume→Latest Writing
2026-09-17
Fused SwiGLU: Tied on Speed, Two Tensors Instead of Four
9th kernel: a fused SwiGLU MLP. Level with torch.compile on time, but holding two big activation tensors between forward and backward where eager holds four — because a compiler will fuse for you, and won't throw an activation away.
2026-09-12
The Leaderboard Changes Every Week, the Excitement Never Outlives Criticisms
Excitement arriving in an afternoon lasts only as long as a frontier model holds its ranking — one or two weeks on average, and you never test the model yourself.
Open Discussion
Loading…
Visitor FAQ
Loading…