6+ years across R&D, application development and research. Today I build agentic AI platforms for Taiwan national agencies: MCP servers, tool-calling analysts that cite every source, RAG retrieval, forecasting models and dashboards, shipped as tested, containerized FastAPI services with CI/CD. Underneath sits real applied-ML and GPU depth (PyTorch, CUDA, NVIDIA PhysicsNeMo) from peer-reviewed research, which most LLM engineers don't have.
Multi-agent and RAG applications with LangChain / LangGraph and AutoGen. Tool-calling loops, MCP client and server, hybrid retrieval (FAISS + BM25 fused by reciprocal rank), and prompt engineering across OpenAI, Claude, Groq, Mistral, Hugging Face, NVIDIA-hosted and Ollama, with the LLM wrapped around a deterministic, cited, tested core. Shipped a seven-role research crew over one shared workspace, gated on an approved plan.
Systems that ship and survive: FastAPI services, Docker, GitHub Actions CI/CD, benchmark-validated test suites, PyPI packaging, model cards, and live demos. Built so anyone can reproduce the headline number themselves.
A real GPU and modeling foundation from peer-reviewed research: PyTorch, CUDA, NVIDIA PhysicsNeMo, neural operators, time-series forecasting, and uncertainty quantification. This is the depth behind the engineering.
Taiwan Polar Institute · Taichung, Taiwan
POUK YAM · agentic AI and automation consultancy · Ouagadougou, Burkina Faso
AI Smart Innovation Foundry (ASUS), Feng Chia University · Taichung, Taiwan
Institute of Applied Geology, National Central University · Taoyuan, Taiwan
AnCAD Co., Ltd. · New Taipei City, Taiwan
Feng Chia University · Taichung, Taiwan
The open-source platform behind my commissioned work at the Taiwan Polar Institute: water data, hydrology, and agricultural water management in one package. 37 unified data sources (USGS, Brazil ANA, Australia BOM, FAO, GEMStat, EU WFD and more), Bulletin 17C flood frequency, FAO-56 crop water, baseflow separation, and an agentic AI engine that scores 26 research methodologies, checks every step against a method-precondition registry, and auto-executes the pipeline. On top of that sits Studio, a seven-role research crew that takes a problem in plain language and returns a finished deliverable bundle. CAMELS-validated, released on PyPI and citable by DOI.
The work I led as principal developer at ASUS / Feng Chia University, open-sourced
afterwards. GPU physics-informed neural operators for groundwater: a single
attribute-conditioned operator that generalizes across sites in place of
per-site hand-calibrated models. A real PhysicsNeMo port with .mdlus
checkpointing, a CUDA GPU benchmark, leave-one-well-out generalization, and a probabilistic
forecast with calibrated intervals. It also reports where it still trails the gray-box
baseline.
photomy tilapia are gasping at the surface this morning
vision: fish at surface, rapid gill movement · guard: no numbers, no prescriptions
A personal agronomy agent for aquaponics, serving 45+ farmers in central Taiwan across 2,000+ consultations. Instead of retrieving what a paper said, Agronaut computes the answer for your specific system: a tool-calling LLM collects facts and routes, while a deterministic, fully tested engineering core does the math, calibrated to your own measured results within published bounds. It remembers your system across sessions and reasons over a curated troubleshooting knowledge base. Every result lists the coefficients it used (with sources) and what it does not model, so you always know the limits of the design.
pip install agronaut is the whole installclaude > /paper-agent review manuscript.docx ✓ runtime Claude Code plugin v1.5.0 ✓ citations Semantic Scholar MCP ✓ guard anti-fabrication: 0 invented refs → manuscript.reviewed.docx
A Claude Code plugin that turns the agent into a disciplined manuscript
collaborator: five modes (draft, review, revise, proofread, audit), citations resolved
through the Semantic Scholar MCP (auto-wired on install), and hard
anti-fabrication guardrails with pause-and-confirm after every section. Calibrated for
hydrology and IEEE journals, with a generic profile for any quantitative-science field and a
clean .docx round-trip.
/plugin marketplace add Rekin226/paper-agent
/plugin install paper-agent@paper-agent
$ make_results_doc.py ASR · held-out test · n=564 MMS-1b mos, fine-tuned 16.8% WER MMS-1b mos, zero-shot 31.1% Whisper-small, tuned 34.1% → blind A/B, translation model: 50% win rate, p = 0.58 no detectable gain
Open translation and speech recognition for Mooré, a language of Burkina Faso with almost no machine-translation or ASR support. I audited what public Mooré data actually existed, built the corpus that was missing, fine-tuned translation and speech models on it, and then ran a blind native-speaker evaluation that contradicted my own automatic metrics. Three models are published on Hugging Face and the repo says plainly where the results do not hold up.
mos adapter, against 31.1% for the zero-shot baseline it started from and 34.1%
for a Whisper-small fine-tune, all on the same 564-item test set$ deploy site-platform --env azure ✓ stack Django · React · PostgreSQL · Azure ✓ modules sites · crews · materials · GIS ✓ status live · commercial client → construction-site management, shipped
Client delivery through POUK YAM, my part-time agentic-AI and automation consultancy: scheduling, crews, materials, and GIS-based site tracking in one Django / React application on Azure REST services. I designed it, built it, deployed it, and I support it in production. Client and product details stay confidential.
6+ years across R&D, application development and research, building production AI: generative and agentic systems, multi-agent and RAG apps served as tested, containerized services, on top of a real applied-ML and GPU foundation. What I bring that most LLM engineers don't is a scientist's discipline. I respect the constraints, validate out of sample, and never ship a number I cannot reproduce.
That discipline is earned, not claimed. During my Ph.D. at Feng Chia University I built physics-informed gray-box models of the Chou-Shui Chi alluvial fan in central Taiwan and predicted groundwater across 33 wells with RMSE between 0.07 and 0.24 m (companion code (opens in new tab)). At ASUS I led the team that replaced those per-site hand-calibrated models with a single neural operator that generalizes across sites. That is the testing rigor I bring to every system I ship.
Today I am Applied AI Engineer at the Taiwan Polar Institute, where I own end-to-end delivery of a multi-year commissioned engagement for Taiwan national agencies, live since August 2025 behind a public government portal, and open-source the platform underneath it. Alongside that I run POUK YAM, a part-time agentic-AI and automation consultancy, and teach numerical analysis and scientific Python as an adjunct instructor at Feng Chia University.
Open to AI/LLM engineer, generative-AI / agentic-systems engineer, and applied-AI / ML engineering roles. Based in Taichung, Taiwan (Taiwan resident, ARC).
4 peer-reviewed papers (3 SCIE-indexed, incl. Hydrogeology Journal 2026) and 1 Taiwan utility model (M661364).
Read on Google Scholar (opens in new tab)Open to AI/LLM engineering and applied-AI roles, collaborations, and consulting.