Rujira Sangswang 王懿阳

AI Application Engineer @ Suzhou TFC Optical Communication Co., Ltd

About

AI Engineering student at Harbin Engineering University with hands-on experience across the full LLM stack — from fine-tuning to production deployment. At TFC Optical Communication, I built an enterprise RAG system now serving 15,000+ employees company-wide: hybrid dense + BM25 retrieval with Chinese-language tokenization, cross-encoder reranking, department-level access control, and a question Ragas evaluation benchmark. I also led a mid-project LLM migration (Qwen → DeepSeek) with full regression testing to confirm zero quality loss. Before that, at Siam.AI Cloud, I worked on domain-specific fine-tuning of Thai language models using NVIDIA NeMo — building SFT/RL data pipelines and training on multi-GPU infrastructure to improve alignment and response quality. Outside of work, I'm building EdgeMechanic, an offline agentic LLM assistant for HVAC technicians (Qwen3 + RAG + LoRA), and I stay active in AI competitions — 1st Runner-Up at FIS × CIP (SinoGo AI Platform), 1st Place in a provincial AIGC competition. Trilingual in English, Mandarin Chinese, and Thai. Currently seeking an AI Engineering internship for Jan-Mar 2027.

Country

China

City

Harbin

Industry

Computer Software

Skill

Reinforcement Learning, Supervised Fine-Tuning, Prompt Design, Docker, Retrieval-Augmented Generation (RAG), Financial Analysis, AI Software Development, Generative AI, Natural Language Processing (NLP), SQL, Microsoft Excel, Chinese, Python (Programming Language), JavaScript, HTML, Machine Learning

Experience

AI Application Engineer

Suzhou TFC Optical Communication Co., Ltd

2026-7 - Present · 3 mos

ซูโจว

Project: Internal Support RAG System for Company-Wide Knowledge Access - Designed and built an end-to-end Retrieval-Augmented Generation (RAG) pipeline for internal IT/HR policy Q&A, scaling to support ~15,000 employees company-wide. - Implemented hybrid retrieval combining dense vector search and BM25 keyword search with Chinese-language tokenization, enhanced with cross-encoder reranking for retrieval precision. - Designed a department-based access control system integrated with the company's employee database, filtering unauthorized content before it reaches the LLM. - Identified and remediated multiple security vulnerabilities across the system, validating each fix through exploit-simulation testing before deployment. - Built a 50-question evaluation benchmark using Ragas (Faithfulness, Context Precision/Recall, Answer Relevancy) and conducted concurrent load testing to validate production readiness. - Migrated the LLM backend from Qwen to DeepSeek mid-project, running a full evaluation re-run to confirm no quality regression.

Artificial Intelligence Engineer

SIAM.AI Cloud

2026-1 - 2026-3 · 3 mos

กรุงเทพมหานคร, ประเทศไทย

Project: Domain-Specific Adaptation of Thai Language Models using Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) with NVIDIA NeMo (NeMo AutoModel & NeMo RL). * Designed and implemented data pipelines for domain-specific Thai datasets, including data preprocessing, curation, and formatting for Supervised Fine-Tuning (SFT). * Conducted model training and optimization using NVIDIA NeMo (NeMo AutoModel and NeMo RL), applying Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to improve model performance and alignment. * Evaluated and analyzed model performance through benchmarking, error analysis, and iterative refinement to enhance domain-specific accuracy and response quality for Thai language tasks.