Thodsaporn Chay-intr

Chief AI Officer at iApp Technology, Bangkok

Portrait of Thodsaporn Chay-intr

I build Thai-language AI at iApp, where I lead the AI and R&D departments and still write code most days. My team's models include OpenThai Legal, OpenThaiGPT, ChindaLLM, and ChindaMT; together with our speech and vision systems, they have served more than 85 million API calls. I work across the whole stack, from post-training with SFT, DPO, and GRPO down to vLLM and TensorRT-LLM serving on our own GPUs.

I hold a doctorate from Institute of Science Tokyo (formerly Tokyo Institute of Technology), where I worked on character-based word segmentation in the Okumura–Funakoshi Lab. I also contribute serving benchmarks to ThaiLLM, Thailand's sovereign language model programme, and consult informally on its inference serving.

The record is public: Tech Collective named iApp a top-five AI startup in Thailand in 2026, our face anti-spoofing holds an iBeta Level 1 certification, and the models and papers are on open.iapp.co.th, openthai.ai, and arXiv.

Download résumé Open models Email me


Work I'd point you to first

OpenThai 2.0 ASR (2026) — Streaming Thai speech recognition on NVIDIA Nemotron, co-built with my OpenThai Lab from 12,373 hours of Thai speech. At 560 ms latency it makes 25% fewer errors than Typhoon ASR on FLEURS (8.47 against 11.28 CER), and one H100 serves about 20 live streams.

OpenThai 2.0 OCR (2026) — Open-weight Thai document readers, co-built with my OpenThai Lab. OCR-Markdown (0.63B) scores 95.2 on Thai document tables against 91.6 for Typhoon OCR 2B, a model three times its size, and drew more than 9,400 downloads in its first five days. OCR-Text is a 1M-parameter model that runs on a CPU at 0.39 seconds a page and beats Google Cloud Vision on printed Thai, 87.8 against 83.6.

OpenThai 2.0 Legal (2026) — I led OpenThai 2.0 Legal, an open-weight Thai legal language model we built with NVIDIA, BDI, and AIEAT on Nemotron-3-Nano-30B. Legal Thai is formal, precise, and badly served by general models; this one is free to download and run, and it was selected for TH-AI Passport, the national AI programme targeting five million Thais.

CIBbot (2026) — I led CIBbot, a legal-advice and incident-reporting assistant we built for the Royal Thai Police's Central Investigation Bureau. It runs fully air-gapped over seven bureau archives and answers with 96.6% retrieval accuracy on a 208-item legal evaluation. It was launched at CIB headquarters with the National Research Council, and you can try it.

open.iapp.co.th and openthai.ai (2026) — I led open.iapp.co.th, which gives iApp's models away under Apache-2.0, including Chinda LLM 4B, ChindaMT, and the ChindaTTS voices. The OpenThai family lives on openthai.ai, launched with AIEAT: ten open models from 1M to 72B parameters, from LLMs to OCR and speech.

ThaiLLM (2025–) — I contribute vLLM and TensorRT-LLM serving benchmarks on H100 to Thailand's national 8B and 30B model programme, and consult informally on its inference serving.

OpenThaiGPT 1.6 & R1 (2025) — Thai-centric reasoning models I contributed to as part of the team. R1 is built on DeepSeek-R1-Distill-32B and beats DeepSeek-R1-70B on Thai reasoning benchmarks at half the size: 56.7 against 33.3 on AIME24-TH. There is a paper, and the weights are public: 1.6-72B, R1-32B.

Certified biometric liveness (2024) — I led our dual-stage face anti-spoofing work, which passed iBeta Level 1 Presentation Attack Detection testing with a 0% attack presentation classification error rate. We have since extended it to passive deepfake screening for banking.


News


Selected publications

The full list, with code links, is on the publications page.