Trần Đại Nhân

• Capstone: S.M.I.L.E, a multi-clinic dental practice management platform. Owned appointment booking, payments and identity verification across a microservice backend.
• Coursework: distributed systems, database systems, software architecture, machine learning, computer vision, data structures and algorithms.
• Produced two accepted conference papers, founded an AI startup and led the university IT club alongside full-time study.

• Led a team building a multi-agent system to validate and reproduce baseline papers, parsing paper claims, repository artifacts and GitHub issue threads into a single reconciled completeness checklist.
• Designed the four-stage pipeline (Paper Agent, Repo Agent, Issue Agent, Reconciler) and scoped agent authority to collection and diagnosis only, keeping code generation and final judgment with the human reviewer.

• Run the experimentation workflow for Vietnamese VQA and NLP research: dataset curation, training runs, evaluation harnesses and result reporting for the team.
• Mentor learners through hands-on labs in NLP and multimodal AI, translating current research into exercises engineers can reproduce.
• Author lecture slides, lab notebooks and coding exercises for AI training cohorts, and review assessment material for accuracy.

• Built backend modules for a course management platform in Java and Spring Boot over PostgreSQL, adding Redis caching and Docker containerisation to improve response times and deployment consistency across Agile sprints.

• PyTorch
• Transformers
• vLLM
• Unsloth
• Gemini 2.5

• Designed a modular Vietnamese VQA framework covering dense image captioning, multi-LLM inference across GPT-OSS, Qwen3 and DeepSeek, batch processing and automated evaluation.
• Built a difficulty-aware router that scales inference compute per question across low, high and adaptive modes, cutting end-to-end inference from 3,317 s to 815 s, a 75 percent reduction against the high-compute baseline.
• Reached BLEU@4 0.142, ROUGE-L 0.654 and METEOR 0.421 on ViVQA-X at a quarter of the baseline compute budget.

DetailsCode

• LLaMA-Factory
• MS-SWIFT
• GPT-OSS
• Llama 3.2

• Built the distillation pipeline curating low, medium and high budget reasoning traces from a GPT-OSS teacher for a Llama 3.2 3B student.
• Added final-answer verification and length-based filtering to strip noisy and over-long chain-of-thought supervision, prioritising sample quality over volume.
• Ran SFT and GRPO workflows reaching GSM8K 0-shot 0.8006 and MATH500 0-shot 0.4760.

DetailsCode

• SigLIP2
• MLP projector
• Llama 3.2 1B Instruct
• PyTorch
• Accelerate
• Hugging Face

• Implemented a two-stage LLaVA-style curriculum: stage 1 freezes the SigLIP2 vision encoder and the Llama decoder and trains only the MLP projector for modality alignment; stage 2 unfreezes the decoder for Vietnamese visual instruction tuning on scraped data and data generated with the OpenAI Batch API.
• Ran distributed training with Accelerate using checkpoint and resume handling so long runs survive interruption, then closed the loop with benchmark evaluation and working demos for captioning and visual instruction chat.

DetailsCode

• PyTorch
• Transformers
• FastAPI
• Docker
• Next.js

• Trained a Vietnamese subword tokenizer from scratch rather than reusing an English vocabulary, cutting token fragmentation on Vietnamese text and improving the effective context budget.
• Curated and deduplicated a mixed news and Wikipedia corpus into roughly 2.64B training tokens across pretraining and continued pretraining phases.
• Adapted the model to a narrow five-word quatrain corpus as a controlled study of how far continued pretraining shifts generation style, then packaged it for serving as a Docker-based trainer with a FastAPI backend and a Next.js chat interface.

DetailsCode

• Directed club operations, coding competitions and technical training; ranked 1st among academic clubs at FPT University in 3 of 4 terms.
• Proposed and led project initiatives presented to the leadership board of FPT University's academic clubs union.

• Led a student research team through problem framing, experiment planning and paper writing, delivering an image labelling pipeline for breast cancer data and an optimised clinical prediction workflow across 2 accepted papers.