parinzee / parin

## Model description (bio)

Parin is an AI undergrad at Carnegie Mellon’s School of Computer Science and an undergraduate researcher at CMU NeuLab, advised by Prof. Graham Neubig. He studies on a full scholarship as an SCBX Young Tech Scholar.

His interests center on making it cheaper to teach LLMs and agents complex tasks: synthetic data and environments, curricula, and reward design, especially when compute, data, or human supervision is scarce.

His most recent paper, Think Before You Link (EMNLP 2026, with Simran Khanuja and Graham Neubig), studies multilingual entity linking: connecting a name in a caption or a face in a photo to the right Wikipedia entry, even when the entity is rare. It shows that state-of-the-art accuracy drops 15.4–39.9% on rare entities, and that a training-free model which searches Wikipedia while it reasons beats the previous best by 6.9% overall and by up to 23.3% on rare entities.

He started doing research in high school. At VISTEC, advised by Dr. Peerat Limkonchotiwat and Prof. Sarana Nutanong, he built a seed-free framework for generating Thai instruction-tuning data (ACL SRW 2024): a curated 5K-example subset beat training on all 100K, and two national AI programs, WangchanX and AI Singapore, adopted the data.

In 11th grade he created Typhoon Vision at SCB10X, Thailand’s first vision-language model, end-to-end as a Multimodal Junior Scientist. He is also a co-author on Typhoon 2, an open family of Thai text and multimodal LLMs released in December 2024.

Outside research, he interned as a data scientist at Siam Commercial Bank, building a causal uplift model, and built self-verifying LLM agents at Cognitus (since acquired by IBM). He TAs 15-213, CMU’s Introduction to Computer Systems.

## Training data (work experience)

6 stops so far, newest first.

• CMU NeuLabAug 2025 – now Undergraduate Researchernow Research on teaching LLMs and agents complex tasks with less compute and supervision, advised by Prof. Graham Neubig. First-author paper at EMNLP 2026 on multilingual entity linking.
• CMUFall 2026 Teaching Assistant, 15-213now Introduction to Computer Systems: office hours and recitations.
• Siam Commercial BankJun – Jul 2026 Data Science Intern Built a causal uplift model that ranks customers by how much a campaign changes what they do, not just how likely they are to buy. In offline evaluation (Jan–Feb 2026 data) it was estimated to deliver 16% more conversions at the same send budget than the bank’s existing ranking.
• Cognitus (acq. by IBM)2024 – 2025 AI/ML Engineer Built self-verifying LLM agents (planner, executor, verifier) that author workflows across 100+ custom node types on the company’s workflow platform.
• SCB10XMar 2024 – Jan 2025 Multimodal Junior Scientist Led and built Typhoon Vision (8B), Thailand’s first vision-language model: data pipeline, training, and release, with support from the Typhoon team. At release it beat Qwen2-VL and Llama-3.2 on Thai visual question answering. Co-author on Typhoon 2.
• VISTECMay – Aug 2024 Synthetic Data Research Intern Generated 100K+ Thai instruction-tuning examples grounded in Wikipedia; a curated 5K subset beat training on all 100K. Adopted by two national AI programs, WangchanX and AI Singapore.

## Awards

• 2026 Heidelberg Laureate Forum Thailand’s sole computer-science representative, nominated by NSTDA; one of 200 young researchers worldwide
• 2026 HackIllinois three awards for OpenReality: First Place, Best AI Inference (Modal track); MLH Best Use of Google Gemini API; Most Creative
• 2025 SCBX Young Tech Scholar full scholarship; sole recipient from Thailand in the 2025 cohort
• 2024 AI Thailand Hackathon 1st Runner-Up with a multimodal Thai tourism agent, later the MAHASAMUT paper
• 2023 SCB10X Bangkok AI Hack 1st place (Track 1), plus the VISTEC and AWS bounties
• 2023 Super AI Engineer, Season 3 Gold Medal, the program’s top distinction; one of seven Gold Medalists, and the first high schooler to win it
• 2022 AI Builders 2 1st place and Best NLP project

## Other projects

• 2026 Resnik CMU NASA SUITS team Built the team’s end-to-end AI voice assistant, designed to support astronauts during spacewalk (EVA) simulations.
• 2026 OpenReality HackIllinois 2026 A phone browser streams a room into a live 3D map. I built the autonomous agent that decides what to look for in the space while mapping continues. Three awards.

## Limitations and biases

• Strong prior toward science fiction (Red Rising, To Sleep in a Sea of Stars, Unto Leviathan).
• Currently in marathon training. Progress: slow.
• Occasionally generates fiction novels.
• Biased toward lower-resource languages, by design.
• Trained mostly on limited compute, so it tends to ask how to get more out of less.

## Intended uses

Research collaborations on agents, synthetic data and environments, curricula, and reward design: making it cheaper to teach models complex tasks when compute, data, or supervision is scarce. Still happy to talk multilingual NLP and lower-resource languages like Thai. Also suitable for hackathons, and for conversations about philosophy, science fiction and running.

## Citation

If this model was useful, please cite it:

@misc{pengpun2026parin,
  title        = {parinzee/parin: Model Card for Parinthapat Pengpun},
  author       = {Pengpun, Parinthapat},
  year         = {2026},
  version      = {v2026.10},
  howpublished = {\url{https://parinzee.github.io/about}},
  note         = {Pretrained in southern Thailand; fine-tuned at
                  VISTEC, SCB10X, and CMU NeuLab}
}

Citing the research instead? Think Before You Link is on arXiv, and everything else is on Google Scholar.

## Model description

Off the clock, Parin is from southern Thailand and loves nature, despite spending most of his days at a terminal. He has long loved science fiction, writes fiction now and then, and keeps a blog to record and reflect before moments slip away. He is also training for a marathon.

## Pretraining corpus

Science fiction keeps his sense of wonder alive.

• Red Rising
• To Sleep in a Sea of Stars
• Unto Leviathan

## Current training run

run       marathon-training
status    ● in progress
progress  slow

## Generations

Writes fiction now and then, and a blog to record and reflect before moments slip away.

## Influences

Thinkers he admires:

## Side projects

• SuperAI Season 3 image-processing hackathon entry ↗ ImageProcessingHackathon-SuperAI
• A Gentoo overlay for Microsoft Surface laptops ↗ linux-surface-overlay
• A React Native + Django app for Bangkok Christian International School ↗ BCIS-app