Machine learning & AI engineer — veterinary surgeon turned developer | Bangkok
AI Engineer Intern at AXONS (CP Group), working on AI and NLP engineering.
I came to machine learning from veterinary neurosurgery — DVM and MSc (First-Class Honours) from Chulalongkorn University, eight years of clinical specialty practice, a published implant-biomechanics paper, and two years coordinating bioequivalence studies under ICH-GCP for Thai FDA submissions. What that background gives me is not a shortcut to ML: it is a hard-wired instinct for evaluation discipline — controls, pass bars fixed before the run, uncertainty quantification, and refusing to believe a number whose provenance I cannot trace.
I retrained through Super AI Engineer Season 6 (Artificial Intelligence Association of Thailand) — a seven-month national program covering classical machine learning, computer vision, statistics, and building an LLM from scratch — completed alongside part-time clinical work.
A note on links: most of the projects below live in private repositories. The two public ones are linked; the rest are named and described but not linked, because a 404 helps nobody. Happy to walk through any of them.
equity-return-ranking — Cross-sectional equity return ranking over a frozen 30-stock US panel, with a deterministic paper-only replay. Places no orders and claims no profitability. The point of the project is the evaluation discipline:
- Expanding walk-forward folds with purge and embargo; train-fold-only preprocessing
- One 2025 holdout year, revealed exactly once, after model selection was frozen
- Baselines first (no-signal, momentum), then Elastic Net and histogram gradient boosting
- 2,000-sample moving-block bootstrap, robustness slices, cost sensitivity
- Realistic costs (1 bp fee + 5 bps slippage per side), immutable hash-named evidence packages, 53 tests, CI
cat-fgs-llm — Feline Grimace Scale pain scoring. The headline deliverable is not the classifier but a portable, power-aware confound-attribution protocol for fine-grained animal-affect models — a per-action-unit audit that validates on planted positive and negative controls. Engine is a frozen DINOv2 ViT-S/14 with five per-AU CORN heads. 16 architecture decision records, and a strict gate run-order where no downstream number is believed until its gate passes. The 0–10 severity layer is marked inspected-not-validated and never emits a validated-claim number.
Election OCR — Structured voting data extracted from 846 scanned Thai
election result documents (Form สส.6/1) from the 2026 general election. A
multi-model OCR pipeline (Gemini 3 Flash + 2.5 Pro + 2.5 Flash) with
template-based party-name alignment, fuzzy assembly, and an 8-test local TDD
eval harness. 16 Kaggle submissions across 3 overnight runs, 900+ OCR API
calls, orchestrated with Claude Code /loop for autonomous monitoring.
Parasite Egg Detection — YOLOv11m object-detection model for identifying parasite eggs in microscope smear images, fine-tuned with paper-inspired hyperparameters and augmentation tuned for thick-smear specimens.
From Data to Insight — Multi-branch time-series demand forecasting for a coffee-house chain.
3rd Place, Mental Health Track.
Used MiroFish (a multi-agent swarm intelligence engine) to model whether "vibe coding" produces a developer-demand boom or bust across 2026–2028.
- 168 rounds, 40 agents, 1,484 total actions
- Agents represented startup founders, security engineers, rescue engineers, AI tool companies, regulators, and Thai SMEs
- Each agent carried an independent personality, long-term memory (Zep), and behavioural logic
- Seeded with 50+ real research citations (NYU, Stanford, Veracode, IBM breach reports)
Key finding: developer demand bifurcates — senior security and rescue engineers see surging demand while junior prompt-only developers face commoditization. Not a uniform boom or bust, but a split.
A licensed veterinarian building AI for animal healthcare — a domain with almost no existing AI tooling.
| Project | What it does |
|---|---|
| vetblood-ai | AI-powered animal blood test analyzer for Thai vet clinics — Gemini 3 Hackathon |
| Clinical AI assistant | Deployed at a Thai animal hospital — SOAP generator, CEO dashboard, and LINE chatbot in one stack |
| Veterinary knowledge graph | Graph-first vet knowledge base — replaces a free-text reference with structured, queryable relationships |
| Pet owner member portal | Thai vet hospital member portal — vaccine history, appointments, and loyalty points |
Open-source tooling that makes Claude Code better for everyone.
| Project | What it does |
|---|---|
| claude-code-statusline | Real-time rate limit %, context window %, session cost, and git status in your statusline |
| claude-code-notify | Push notifications to Mac, iPhone, and Apple Watch when Claude Code needs input |
| drawio-mcp | MCP server for creating draw.io diagrams from LLMs — XML, Mermaid, and CSV |
| awesome-claude-skills | Curated list of 100+ Claude Code skills across dev, data, DevOps, and more |
| ai-anti-hallucination | Anti-hallucination rules for AI coding agents |
Reusable patterns extracted from production AI systems I've built.
| Project | What it does |
|---|---|
| soul-agent-framework | Configure AI agents through markdown, not code — SOUL/MEMORY/skills architecture |
| agent-factory | Spin up fully-deployed AI agents for any domain in under 2 minutes |
| agency-orchestrator | Multi-agent tmux orchestrator for Claude Code — 167 specialized agents in visible parallel waves (Research → Build → QA) |
| arena-workflow | Competitive multi-model dispatch — every task gets every model, evidence picks the winner |
| openclaw-actual-budget | AI agent template — receipt scanning and expense tracking via Telegram + Actual Budget |
| obsidian-ai-knowledge-agent | AI agent for Obsidian — Second Brain, Personal Search Engine, AI Coding Vault, Karpathy LLM Wiki |
I run a small Thai SME web agency on the side — 20+ sites shipped across 7 industries (vet clinics, industrial B2B, hospitality, services, construction), and I'm consolidating that work into a multi-tenant CMS. It is where I learned to ship and maintain software for people who will never read a changelog.
- DVM and MSc Veterinary Surgery (First-Class Honours, GPA 3.95) — Chulalongkorn University
- Former veterinary neurosurgeon — spinal-injury and epilepsy caseload; published implant-fatigue biomechanics research
- Medical writer & clinical research coordinator — bioequivalence studies under ICH-GCP for Thai FDA submissions
- Super AI Engineer S6 (AIAT) — seven-month national AI program
- I write about ML, local LLMs, and agent design at mingrath.com