chinavat ~ % whoami
Chinavat
I'm a passionate developer and ML engineer focused on innovation, with experience across AI, games, and robotics.
- % education
- B.Sc. Applied Computer Science
- King Mongkut's University of Technology Thonburi (KMUTT)
- Graduated 2026 · GPAX 3.69
- Petchra Pra Jom Klao scholarship.
% find portfolio
→ showing all 17
% ls projects/(6)
Things I've built — ML systems, apps, and side experiments.
NeuroGami
A creature-battler where you don't grind stats — you train living neural networks.
#reinforcement-learning #machine-learning #game-dev
★2026A creature-battler where you don't grind stats — you train living neural networks. Every monster carries its own machine-learning brain. There are no scripted move-sets: each fight, each decision, and each outcome becomes training data, so monsters genuinely get smarter the more they battle. Strategy means shaping a learning agent rather than farming numbers — a playable demonstration of online, experience-driven learning wrapped in a familiar monster-training loop.
- year
- 2026
- featured
- true
- status
- Live-service title in active development — KMUTT final project, continued post-graduation (private repo)
Makes reinforcement learning the core game mechanic: the player is effectively curating a training curriculum. End-to-end execution — ML model design, real-time inference inside a game loop, and design that makes the learning legible and fun.
Brainboost
Turns any class lecture into a microlearning game you can finish on the bus.
#llm #mobile #edtech #gamification
★2025Turns any class lecture into a microlearning game you can finish on the bus. A Flutter mobile app that ingests lecture material and uses an LLM backend to automatically generate bite-sized, game-like learning content from it. Instead of re-reading slides, students review through short interactive rounds — lowering the activation energy of studying and turning passive recordings into active recall. Built end to end: mobile client, LLM content-generation pipeline, and the game loop that ties them together.
- year
- 2025
- featured
- true
A real LLM product, not a demo: a generation pipeline feeding a polished Flutter app. Demonstrates applied GenAI, mobile development, and a sharp read on the learning-science problem (microlearning, active recall).
GRPO-Based Reinforcement Learning for Flow-Matching VLA Models
An experiment using GRPO to teach a SmolVLA robot policy to improve itself through trial and error.
#reinforcement-learning #robotics #vla #research
2026An experiment using GRPO to teach a SmolVLA robot policy to improve itself through trial and error. A class research project that adapts GRPO — the reinforcement-learning algorithm behind today's reasoning models — to SmolVLA, a flow-matching vision-language-action model. The policy acts inside a simulated manipulation environment, and the reward from each rollout is meant to push behavior past what supervised imitation alone can reach. Marked as a failed experiment: training was bottlenecked by CPU-only simulation, which made rollouts too slow for the policy to converge.
- year
- 2026
An experimental pairing of GRPO (DeepSeek-style RL) with a flow-matching VLA policy — still rare in published robotics work. Touches the full embodied-AI stack: policy modeling, simulation rollouts, and reward-driven fine-tuning.
MADX — Multiplayer AI Diffusion eXperience
A multiplayer diffusion game engine where the player actually plays inside the model, not a program.
#diffusion-models #reinforcement-learning #world-models #game-dev
2025A multiplayer diffusion game engine where the player actually plays inside the model, not a program. An experiment inspired by GameNGen (DOOM running inside a diffusion model). MADX is a diffusion-based world model trained on trajectories collected from multiple agents interacting in an environment (currently ATARI Boxing). Instead of rendering a hand-coded game, the model generates the next frame conditioned on both players' actions, turning a learned simulator into a real, two-player game — the game world is the neural network itself. Ships with public code.
- year
- 2025
A working, playable world model rather than an offline benchmark. Brings together diffusion modeling, multi-agent data collection, and real-time interactive generation, with open-source code on GitHub.
Gamify Habits Tracker
A habit tracker disguised as a monster-catching game.
#mobile #gamification #game-dev #product-design
2024A habit tracker disguised as a monster-catching game. Monsters spawn at random, and the only way to feed and tame them is to complete your real-world tasks. Each finished habit earns food; enough food tames the monster; and the "catch them all" collection loop keeps you coming back tomorrow. It reframes the dull discipline of habit-building as a light, rewarding game, borrowing collection and progression psychology to make consistency actually stick. All in-game assets are produced through a pixel-art generation pipeline.
- year
- 2024
Product thinking first: applies proven game-design reward loops — variable rewards, collection, progression — to a behavior-change problem, taken from concept all the way to a working app.
Payex — Payment Extraction
Reads a messy 7-Eleven receipt and returns clean, structured line items.
#computer-vision #multimodal #ocr #llm
2024Reads a messy 7-Eleven receipt and returns clean, structured line items. A computer-vision pipeline that extracts itemized information from convenience-store receipts by chaining several specialized models: Grounded-SAM for region detection, CRAFT for text localization, TrOCR for recognition, and an LLM to normalize the raw output into structured data. A practical document-understanding system assembled from modern multi-modal building blocks instead of forcing a single black box to do everything. Built at a time when ChatGPT still couldn't read text from images — so the pipeline had to do all the visual reading itself.
- year
- 2024
Real-world document AI: composes four models (Grounded-SAM, CRAFT, TrOCR, LLM) into one robust pipeline. Shows systems thinking — picking and stitching the right tool at each stage of a noisy, real-input problem.
% ls researchs/(1)# research
Papers and academic work.
Google Play Review Quality Scoring for Digital Engagement and App Development Using Transformer Models
Scoring how useful an app-store review is — not just whether it's positive. Peer-reviewed and presented.
#nlp #transformers #research
2025Scoring how useful an app-store review is — not just whether it's positive. Peer-reviewed and presented. A transformer-based NLP study that moves beyond positive/negative sentiment to estimate how informative a Google Play review is for driving digital engagement and guiding app development. The work was accepted and presented at iSAI-NLP 2025. It connects raw user feedback to concrete product decisions, treating review text as a signal for what to build next.
- published
- 12–14 November 2025, iSAI-NLP
Peer-reviewed and presented at iSAI-NLP 2025. Covers the full research arc — problem framing, transformer modeling, evaluation, and publication — with a results-oriented framing aimed squarely at product impact.
% ls jams/(2)# game jams
Games shipped against a 48-hour clock.
Dungeon Delivery
A top-down adventure with a twist: the longer you play, the weaker you get.
#game-jam #game-dev #action
2023A top-down adventure with a twist: the longer you play, the weaker you get. Built for Global Game Jam 2023. You're a cursed knight tasked with returning a player's equipment to the dungeon. You start almost unstoppable, but your power drains as you progress — inverting the usual power-fantasy curve and forcing you to spend your early strength wisely before it's gone. One clear mechanical hook, delivered in a 48-hour sprint. We didn't take it any further — a small team with only a couple of artists, so it stayed at the jam build.
- jams
- Global Game Jam 2023
An inverted progression curve as the central design idea, executed end to end in a 48-hour jam. Shows hook-first game design and the ability to ship a coherent experience under pressure.
Agnetic
A magnet-packing puzzle built in 48 hours for GMTK Game Jam 2021.
#game-jam #game-dev #puzzle
2021A magnet-packing puzzle built in 48 hours for GMTK Game Jam 2021. Players fit magnets into a box under polarity-style constraints: some magnets repel and can't sit beside each other, others attract and must be adjacent. The challenge is finding the one arrangement that satisfies every rule — easy to pick up, deceptively tricky to solve. A tight, single-mechanic puzzle shipped against the jam clock. It was my first solo game jam against the 48-hour clock — and I kept iterating on this version for about five months afterward, before starting university.
- jams
- GMTK Game Jam 2021
A complete, rule-driven puzzle game delivered within a 48-hour jam — constraint design, level logic, and the scope discipline to finish something playable under a hard deadline.
% cd
Places I've worked.
2024 – 2026
AI Engineer · Computer Vision @ DeepCapital
Shipped production computer-vision systems for factories and large-scale CCTV operations. The computer-vision specialist on the team, taking camera and video problems from prototype to deployment. Built factory automation pipelines (automatic counting and inspection), multi-camera CCTV surveillance, and CCTV optimization — turning raw video feeds into reliable, real-time signals the business could act on.
Real-world CV in production, not benchmarks: factory automation, auto-counting, and multi-camera CCTV surveillance and optimization, owned end to end.
2024
Gold Medalist (top 10 nationally) @ Super AI Engineer Season 4
Won a Gold Medal — one of only 10 awarded — in Thailand's flagship national AI talent program. Super AI Engineer is AIAT's intensive program to develop the country's frontline AI engineers. Medals are decided on total score across three levels: an entry gate (academic test, hackathon, and interview); an immersive training camp with weekly hackathons on real company problems plus a 24-hour hackathon spanning five domains — NLP, data science, signal processing, IoT, and image processing; and on-site deployment at partner organizations. Competed from House Pangpuriye and finished in the national top 10.
Gold Medal + certificate + 50,000 THB, presented at AI Thailand Forum 2024 (25 Oct 2024). Backed by AIAT, NECTEC, KBTG, and Thailand's Ministry of Higher Education, Science, Research and Innovation.
2021 – 2023
Machine Learning Engineer @ Looloo Technology
Built OCR and text-detection data pipelines end to end, part-time through university. Ran full data pipelines from collection to model evaluation. Spearheaded the Universal Text Detector — owning data engineering, pseudo-labeling, and model-synthesis pipelines for both printed and handwritten documents — and ran QA for the TH–EN Handwritten OCR project, systematically analyzing, tracking, and visualizing model error to push accuracy up.
First industry ML role, held part-time while studying. Led the Universal Text Detector pipeline and the QA / error-analysis workflow for Thai–English handwritten OCR.