Now · Specialist (Lead), AI Solution & Platform at True Digital Group
AI Solution Lead · Research → Production · On-prem AI · AI Products
I turn AI research into production systems that create real value, run at scale, stay on-premise and that people actually use.
What matters is the value AI creates. I take ideas from research to architecture to production rollout — at scale, or fully on-premise where data can't leave — with LLMs, multi-agent systems and voice as the tools.
Worked & published with
01What I do
Three things I'm hired for — each backed by systems running in production today.
I own the whole system, from architecture to production rollout: speech in, agents and retrieval in the middle, speech out — plus the evaluation that keeps it reliable.
I work with product teams to decide what the AI should do, then build it and measure it with real users — not demos.
Private models where data cannot leave the building — quantized, batched and sized to run on the hardware you already have.
Evaluation · observability · FinOps — every stage measured in production
02New research
My newest work: an open method and tool that reads a model's architecture from its file header.
AI Expo Thailand 2026 · Squeeze Until It Breaks: A Hands-On Edge Inference Contest · Sat 7 Nov 2026, 14:00–15:00 · Builder Stage
A free, open-source tool that opens any Hugging Face model and shows its parts, sizes and precision from the safetensors header alone, without downloading the weights. Runs in the browser, as a command-line tool and as a Python package. I built it for my workshop, then turned the method into a paper: most of a model's architecture can be recovered from the tensor names and shapes in a file header of a few tens of kilobytes. Applied to 12,974 Hugging Face repositories, it also shows how open models have changed: mixtures of experts went from 0% of popular language models created up to 2022 to 42% of 2026 releases.
pip install ai-model-explorer
03Highlights
Production AI systems I designed or led — measured by what they do for real users.
TrueX · True Digital Group
800Kmonthly active users
Security compliance zone · 6 sites, Thailand
2,300call-hours per day on one RTX Pro 6000
KBank · KBTG
80%of customer chat traffic automated
KBank · KBTG
93%of call-center use cases covered
04Experience
Telco, banking, research labs and an AI-video startup — from Thai NLP research to leading production voice AI.
True Digital Group
Bangkok, Thailand
May 2025 – Present
Technical AI lead for production voice and LLM systems on the TrueX platform.
KASIKORN Business-Technology Group (KBTG)
Bangkok, Thailand
Feb 2022 – Apr 2025
Research-to-production NLP, speech and LLM systems for banking.
National Electronics and Computer Technology Center (NECTEC)
Pathum Thani, Thailand
Jun 2020 – Dec 2020 · Jul 2021 – Dec 2021
Thai speech and NLP research in the Speech and Text Understanding lab.
Synthesia
London, United Kingdom · remote
Oct 2023 – Jul 2024
Research on AI-driven video synthesis and multimodal systems with international teams.
Thailand Mental Health Tech Discovery, Mahidol University
Thailand
Mar 2021 – Dec 2024
Led a team of 7 building AI health-tech that combines psychology and machine learning.
Siam Commercial Bank (SCB)
Bangkok, Thailand
Jan 2020 – Jul 2020 · Jun 2021 – Nov 2021
Speech emotion recognition for call-center applications.
Faculty of IT, KMITL — PRIP Lab
Bangkok, Thailand
Aug 2019 – Jan 2025
Computational linguistics and pattern-recognition research.
Dolby Laboratories
Australia · remote
Jan 2021 – Jun 2021
Deep learning models for speech and audio processing in an international R&D team (Dolby U).
KBTG — Super AI Engineer program
Bangkok, Thailand
Mar 2021 – May 2021
Deployed a QA search-engine service with Elasticsearch, PyTorch and TensorFlow.
KBTG — Machine Learning Bootcamp
Bangkok, Thailand
Jun 2019 – Aug 2019
Thai dependency parsing with EDU segmentation for 'AI for Thai'. Best Progress & Popular Vote awards.
Wisesight
Bangkok, Thailand
Jun 2018 – Mar 2019
Social-listening analytics with ZocialEye, Python and Tableau.
Village Farm Winery
Thailand
Jan 2018 – Jun 2018
Digitised paper records into a cloud ERP with Python and OCR; customer segmentation.
05Worldwide
Seven countries so far: research and product work with teams in the US, UK, Singapore, India and Australia, competitions and exchanges across Asia, and papers at ACM, IEEE and Springer venues.
ThailandBangkok
United States
United KingdomLondon
Singapore
India
Australia
JapanKurume
06About
AI engineer with 6+ years of experience building production-grade AI products with LLMs, NLP, voice, multi-agent and on-premise AI systems. I design and deploy end-to-end systems that serve large user bases — conversational AI, voice assistants and intelligent automation — and turn research into products with measurable user and business impact.
สัตยา สิงห์กุล — วิศวกร AI จากกรุงเทพฯ ประเทศไทย เชี่ยวชาญการสร้าง AI แบบ end-to-end ตั้งแต่ออกแบบสถาปัตยกรรมจนขึ้น production ทั้ง AI product, ระบบ AI on-premise, LLM, Voice AI, NLP และ multi-agent ให้บริการผู้ใช้จำนวนมาก ตั้งแต่ผู้ช่วยเสียง แชตบอต ไปจนถึงงานวิจัยด้าน speech และภาษาไทย
Thai — NativeEnglish — CEFR B1 · professional working proficiency
LLM / GenAI
AI / ML
GenOps
Programming
Cloud
07Research
Speech emotion recognition, Thai speech and language, wake-word detection and spoken question answering.
2026
Preprint · Method behind AI Model Explorer and the Hub Weight Census
2025
2024
Conference · Presenter · Rank B · HSRI-funded
2023
Conference · Presenter · Best Paper session
08Blog
Lessons from building voice, LLM and on-premise AI systems in production.
Model anatomy from tensor names and shapes alone: 96.8% role accuracy on 510 architectures, a census of 12,974 Hub repos, and AI Model Explorer as a web app and Python package.
Speech emotion recognition looks solved on benchmarks and breaks on real calls. A researcher's tour of domain shift, speaker leakage, metric learning and self-supervised features.
An open-source tool that turns concurrency and model choices for STT → LLM → TTS into VRAM, RAM, CPU, TFLOPS and cloud or on-prem cost.
09Open source
See any Hugging Face model part by part without downloading it: architecture, parameters and precision from the safetensors header. Web app, CLI and Python package (pip install ai-model-explorer), plus a census of 12,974 Hub repositories.
Capacity planning and FinOps estimation for conversational-AI pipelines: VRAM, latency budgets and GPU cost for STT, LLM, TTS and RAG workloads.
Evaluation scripts for Thai text-to-speech, focused on speaker tone and pronunciation.
LTMMv2 — notes on a self-supervised Thai speech foundation model.
Conv-TasNet speech enhancement on CHiME-4 with ESPnet, used ahead of speech emotion recognition.
Coursework notebooks on vector databases and semantic search.
Biaffine dependency parser for Thai on Universal Dependencies treebanks.
10Education & awards
2019 – 2025
King Mongkut's Institute of Technology Ladkrabang (KMITL)
GPA 4.00
Thesis: Verify-to-Classify: A Separation Learning for Generalized Speech Emotion Recognition
2016 – 2020
King Mongkut's Institute of Technology Ladkrabang (KMITL)
GPA 3.74
11Community
IEEE Access, ACM TALLIP, IJEECS · IEEE ICITEE 2025, Springer ICONIP 2025, ACM ICMR 2024
Reviewing NLP, audio and vision research.
2021 – present
AIAT — Super AI Engineer, seasons 2–5
Mentored teams on Thai NLP, speech and LLMs; several went on to win hackathons and team awards.
2022 – 2025
Huachiew Chalermprakiet University
AI futures, speech recognition, image processing and AI frameworks for undergraduates.
2022 – 2025
IT Camp #14, KMITL
Planned the camp and taught data preparation and modelling.
2018
ToBeIT60, KMITL
Taught IT topics to ~500 high-school students.
2016
Kurume National College of Technology, Japan (SAKURA Exchange)
Short-term science research exchange.
2017
Microsoft Thailand
Advised customers on Microsoft products.
2017
11 / Contact
Open to conversations worldwide on end-to-end AI products, on-premise LLM and voice systems, research collaboration and speaking. Email is the fastest way to reach me.
joeysattaya@hotmail.com
