Tran Nam Nhat Anh

<p>Built an end-to-end data preprocessing pipeline that OCR-processed approximately 1M biomedical papers from Semantic Scholar, extracted multimodal content, filtered low-quality Bio/Chem text with LLM scoring, and synthesized over 2M high-quality QA pairs for VLM training.<br>
LoRA fine-tuned Qwen-2.5-VL-32B-Instruct on synthetic biomedical data, achieving more than 1% accuracy gains on Human Last Exam (Bio/Chem).<br>
Performed benchmarking of Qwen-VL (7B/32B/72B) and MedGemma-4B on HLE, MMLU, BioInfoBench, PubMedQA, and related tasks.<br>
Developed a clustering system for bio-informatics plots on approximately 500k biomedical image-caption pairs using BioMedCLIP and LLM.<br>
Designed automated task runners for AutoML System on BioML Benchmark, including biomedical imaging, drug discovery, protein engineering, and single-cell omics, achieving 80–90th percentile leaderboard performance within 24 hours and outperforming all baseline agents.<br>
Fine-tuned Qwen-3-8B-Instruct as a Reward Model using approximately 50k OpenReview papers to automate research idea evaluation and improve AutoML system performance.</p>

Built data preprocessing pipeline for biomedical papers, Fine-tuned models Qwen-2.5-VL-32B-Instruct, Qwen-3-8B-Instruct, Developed clustering system for Bio-informatics plots, Designed automated task runners for AutoML System, Automated research idea evaluation with Qwen-3-8B-Instruct

<p>Developed a modular DataLoader library to automate the downloading, parsing, and preprocessing of major recommendation datasets, streamlining the research pipeline.<br>
Enhanced NAML by 2.3% AUC and 3.1% nDCG@10 using custom abstract embeddings.<br>
Boosted NRMS by 0.4% nDCG@10 via concatenated multi-field context.<br>
Investigated positional encoding for LRU4Rec, achieving 6.7% gains in Recall@1, MRR@1, and nDCG@1.<br>
Enriched ml-100k using hierarchical LLM prompts to generate titles, entities, and abstracts.<br>
Used VLM to extract visual keywords and descriptions from movie posters.<br>
Trained SASRec on fused embeddings, achieving 2.0% nDCG@10 over baseline with entity-enhanced features.</p>

<p>• Developed a modular DataLoader library to automate the downloading, parsing, and preprocessing of major recommendation datasets, streamlining the research pipeline.</p><p>• Enhanced two sequential recommendation baseline: (i) improve NAML by +2.3% AUC and +3.1% nDCG@10 using custom abstract embeddings, and (ii) boosted NRMS by +0.4% nDCG@10 via concatenated multi-field context.</p><p>• Investigated positional encoding for LRU4Rec, achieving +6.7% gains in Recall@1, MRR@1, and nDCG@1.</p><p>• Enriched ml-100k using (i) hierarchical LLM prompts to generate titles, entities, and abstracts, and (ii) VLM to extract visual keywords/descriptions from movie posters; trained SASRec on fused embeddings, achieving +2.0% nDCG@10 over baseline with entity-enhanced features.</p>

<p>Benchmarked 7B LLMs on Vietnamese tasks, including Math, Chain-of-Thought, and MCQ tasks.<br>
Implemented GraphRAG with Neo4j for local recommendation.<br>
Designed an OCR data pipeline featuring Kafka producers and consumers for reliable document streaming.<br>
Integrated the pipeline with Spark and Hadoop for scalable batch processing.</p>

<p>• Benchmarked 7B LLMs on Vietnamese task, including Math, Chain-of-Thought, and MCQ tasks; implemented GraphRAG with Neo4j for local recommendation.</p><p>• Designed an OCR data pipeline featuring Kafka producers and consumers for reliable document streaming, integrated with Spark and Hadoop for scalable batch processing.</p>