Dinh Minh Hai
Mathematics & AI Researcher · Ho Chi Minh City
Ho Chi Minh City University of Education
I am a Mathematics graduate with deep interest in the intersection of Artificial Intelligence and Mathematics. My research spans evaluating LLMs' reasoning on proof problems, recursive statistical estimation, and building intelligent knowledge-retrieval systems. I enjoy rigorous theory and building things that actually work.
Research Interests
AI for Mathematics
LLM mathematical reasoning, redundancy detection in proof problems, and multi-agent verification systems.
Statistical Estimation
Recursive non-parametric estimation, periodic shape-invariant models, and asymptotic convergence theory.
NLP & Retrieval
RAG pipelines, embedding fine-tuning, contrastive learning, and enterprise knowledge-base systems.
Projects & Research
RHIM: Redundant Hypothesis Identification in Math
ResearchHo Chi Minh City University of Education · Oct 2024 – Present
Created a benchmark of 3,439 Proof Problems With Redundant Assumptions (PPWRA) to evaluate LLMs' ability to detect redundancy. Devised logic-based strategies for constructing redundant assumptions and developed a multi-agent system for detection, proof, and automatic verification. Benchmarked GPT-4o, Gemini, o1, and DeepSeek-R1.
🏆 Best accuracy: 49% (o1) — showing leading models still struggle with redundancy identification.
Recursive Estimation in Models of Deformation
ThesisHCMUE Undergraduate Thesis · Sep 2023 – May 2024
Reproduced and extended results from Fraysse (2014). Researched periodic shape-invariant models, the Robbins–Monro algorithm, and Nadaraya–Watson recursive estimation. Provided rigorous proofs of parametric and non-parametric convergence, asymptotic normality, and quadratic laws. Implemented in R with a vectorized algorithm.
🏆 30% execution time reduction via vectorization. Accurately reproduced ECG waveforms for healthy and pathological hearts.
RAG Chatbot for Internal Knowledge Bases
IndustryBig Data Research Center, VinUni · Jul–Nov 2025
Built an end-to-end RAG pipeline on WixQA: multi-strategy chunking, Chroma vector stores, Qwen3-0.6B embeddings. Designed a unified evaluation framework (Recall/HitRate/MRR, BLEU/ROUGE, LLM-judged factuality). Improved retrieval via contrastive fine-tuning + BAAI/bge-reranker. Prototyped hybrid GraphDB+VectorStore RAG with LLM-based entity/relation extraction.
🏆 74.2% Recall@3 · token-based Recall@3: 0.705→0.742 (+5.2%) · MRR@3: 0.487→0.512 (+5.1%)
Sentiment Analysis on Amazon Reviews
IndustryBig Data Research Center, VinUni · Sep–Nov 2025
Multi-class sentiment classification with severe class imbalance. Benchmarked CNN, RoBERTa, DeBERTa. Applied Contrastive Learning for feature representation, and Q-LoRA for memory-efficient LLM fine-tuning.
🏆 10% performance gain over Qwen-0.6B baseline.
Education
Bachelor of Mathematics & Teaching Mathematics
Ho Chi Minh City University of Education (HCMUE)