RuffLogix
Published research in computer graphics, medical imaging AI,
and formal verification — and the systems that put it into production.
About
Software engineer and AI researcher based in Bangkok. Currently full-time at Khui AI on LLM training, fine-tuning, and deployment, alongside an AI engineering apprenticeship at KBTG. Previously at Agoda — bridging LLMs and Temporal.io through MCP — plus LINE MAN Wongnai, AIMET, and Looloo Technology, across federated learning, speech AI, backend systems, and smart contracts.
Computer Engineering student at Chulalongkorn University, publishing on computer graphics, medical imaging AI, and formal verification — 3 papers in IEEE Access, JCSSE, and on arXiv. Super AI Engineer Season 5 · iGEM 2024 Silver · TOI Thailand 2020/2021.
Full backgroundFocus Areas
Selected Work
All projectsA learning management platform where courses are assembled from composable "blocks" — Lego-like pieces (text, video, quizzes, code, and more) that instructors snap together to build lessons. Aimed at flexible, hands-on course authoring.
A self-hostable online judge embedded in the RuffBlitz LMS for programming and algorithms courses. Compiles and runs Python/C/C++ through a sandboxed execution engine (Judge0 → isolate) driven by horizontally-scaled judge workers, with role-based access (admin/staff/student), group-gated problem visibility, ICPC/IOI-style contests with scoreboard freeze, subtasks and testlib checkers. Built on Next.js + Elysia (Bun), PostgreSQL/Drizzle, and Redis/BullMQ.
A lightweight ordinal image classifier that grades urine color on a 5-point scale (Clear → Very Dark) as a coarse hydration proxy, built for CPU inference. Treats the ordered classes as ordinal regression (class-weighted CE + Squared-EMD loss), with color-preserving augmentation, effective-number class balancing, and a held-out + 5-fold stratified protocol. Reaches QWK 0.963 on the test set with 100% adjacent accuracy — every error is off-by-one — while a trivial color+GBM baseline matches ~98% of it, making the efficiency case for tiny deployment. Delivered as a LINE chatbot via the LINE SDK and Messaging API, so users get a hydration reading by sending a photo in chat. Not a diagnostic tool.
Selected Research
All publicationsResearch Philosophy
A result that only runs on the author's machine, on the author's split, is not a result.
Formal models where tests can't reach — concurrent pipelines deadlock in states no unit test enumerates.
Baselines before models. Sometimes copying the last frame beats the network you spent a month training.
The paper and the running system are the same project, finished at different times.
" A proof you can check and an experiment you can rerun —
the same discipline, applied twice.
Writing
All posts(Thanks ChatGPT Image!)เมื่อประมาณ 2 เดือนที่แล้ว ผมได้รับโจทย์มาจากพี่กวง KBTG ให้ทำระบบ “Event Transcription & Real-Time Transla…
ช่วงนี้ ถ้าใครได้เลื่อนดู shorts หรือ reels ก็คงจะเห็นโปรเจคแนว ๆ นี้กันบ้าง ที่เป็นคนเปิดกล้องแล้วทำท่า “six seven” หลังจากนั้นก็…
เนื่องจากวันก่อนผมได้เจอบทความของรุ่นพี่ที่เล่าเกี่ยวกับวิชาเลือกในภาค ซึ่งช่วยผมได้เยอะมากในการตัดสินใจเลือกลงวิชาเลือก… Continue…