I build production-grade systems that scale. My code doesn't just work... it's architected for performance, maintainability, and real-world impact. Shipped software with installs across three desktop platforms, a package on PyPI, live web products with registered users, and peer-reviewed research published at ACM.
I'm obsessed with the details that separate prototypes from production systems. Whether it's optimizing database schemas, building modular API architectures, or deploying ML models at scale, I think deeply about how systems perform under real conditions.
Backend-first thinking. I start with data models, API contracts, and system architecture before touching the frontend. Clean separation of concerns isn't optional... it's how you build systems that last.
ML in production. I don't just train models, I ship them: inference pipelines, batch processing, error handling, and performance optimization for real-time computer vision, speech, and time-series classification.
Evidence over assertion. I benchmark before I commit. On MindSync a 15-model evaluation overturned the text classifier already in production; I then shipped the accuracy-per-parameter optimum rather than the top scorer, trading 0.007 macro-F1 for 3.2× fewer parameters because every cold start loads the weights.
| Project | Traction | What it does | Stack |
|---|---|---|---|
| Bug Gym | 130+ installs · 3 platforms | Offline Java exam-practice environment with in-memory compilation, character-level test diffing, and native releases for Windows, macOS, and Linux. | Java 21, JavaFX, RichTextFX, Maven, jpackage |
| Lexo | 1k+ PyPI downloads · 8 releases | Open-source OCR desktop app and CLI for Burmese documents, scanned books, images, and PDFs. One engine behind two interfaces, three export formats. | Python, PySide6, Typer, PyMuPDF, Pydantic, uv |
| MyanTyper | 80+ registered users | Privacy-first typing tutor for the Windows Myanmar Visual-order keyboard, where the front vowel is typed before its consonant. Six-track curriculum, optional cross-device sync. | Next.js 16, React 19, TypeScript, Better Auth, Drizzle, Neon |
| MindSync | 15 models benchmarked · +0.172 macro-F1 | Dual-channel journaling that analyses text and voice at once and flags when they disagree, via Russell's valence-arousal plane. Basis of the IAIT 2026 paper below. | Flutter, Dart, Riverpod, FastAPI, PyTorch, RoBERTa, wav2vec2 |
| CampusLink | 34 contributors · #2 client contributor | KMUTT's campus platform: 14 modules built by 30+ developers. My team owned the auth and course layer every other module integrated through. | React, Express, Prisma, PostgreSQL, JWT, TanStack Query |
| GodView AI | 🏆 Intl. AI Hackathon 2025 winner | Real-time clinical anomaly detection with YOLOv8 inference on camera frames, role-aware access, automated Telegram escalation. | FastAPI, YOLOv8, Supabase, PostgreSQL |
| OSAS Detection System | 961,357 observations | Leakage-aware multimodal transformer research on physiological time-series, with patient-aware splits across a 30-person cohort. | PyTorch, Transformers, NumPy, Pandas |
| thedevbox | Privacy-first developer utilities that process pasted data locally in the browser. | Next.js 16, React 19, TypeScript, Vitest, Biome | |
| SIT Compendium | Structured academic publishing platform with typed content and first-class mathematical notation. | Next.js, TypeScript, React, KaTeX, Docker | |
| SIT Face Recognition | Multithreaded face-recognition and event-photobooth desktop system that releases camera and model resources when idle. | Python, PyQt6, OpenCV, InsightFace, FastAPI |
Cross-Modal Emotion Discrepancy Detection: Detecting When Voice Reveals
Emotions That Text Conceals
Kaung Hset Hein and Jonathan Chan · IAIT 2026 ·
ACM Digital Library
Research across three labs: applied ML and NLP at KMUTT's Innovative Cognitive Computing (IC2) Lab, deep learning and computer vision at KMUTT's Deep Learning Research Lab, and OCR/NLP pipelines for low-resource languages at Language Understanding Lab, Myanmar.
Languages
Frontend & apps
Backend, data & ML
Delivery & tooling
The intersection of elegant architecture and measurable impact. I'm fascinated by how the right database schema transforms query performance, why proper separation of concerns keeps codebases maintainable at scale, how thoughtful error handling turns a prototype into a production system, and what it takes to deploy ML models that hold up outside a notebook.
- International AI Hackathon 2025, 1st place
- CSC 302 Talkathon 2025, Most Creative Topic Award
- The Hackathon 2019 by Ooredoo and UNDP, 1st runner-up
- KMUTT Multi-intellectual Scholarship, full scholarship