AI Research Intern Candidate: Tran Quoc Khanh

TRAN QUOC KHANH

AI Research AI Engineer 091*-***-*** ********************@*****.*** GitHub: Khanhhh239 Ho Chi Minh City EDUCATION

University of Science – VNU Ho Chi Minh City (HCMUS) 2024 – 2027 (expected) Bachelor of Science in Artificial Intelligence (Priority enrollment), GPA : 3.6/4 TECHNICAL SKILLS

Languages: Python (advanced), C++ (intermediate)

Frameworks: PyTorch, TensorFlow, Hugging Face Transformers, NVIDIA NeMo Efficiency: Quantization, Adapter and LoRA fine-tuning Task: Detection, Segmentation, Restoration, Generation, Question Answering and Text-based Retrieval. PROJECTS

STELLAR-RAG — Hybrid RAG Agent for University Documents Python 2026 GitHub: github.com/Khanhhh239/STELLAR_RAG

• Objective: A Vietnamese university Q&A system that answers questions about PDF documents.

• Techniques: Implement a hybrid retrieval mechanism fusing dense (bge-m3 + FAISS), sparse (BM25), and knowledge- graph searches via QDAP-S adaptive fusion. The architecture integrates an EHRAG entity hypergraph with diffusion re-scoring, a HybGRAG critic loop, Self-RAG, and Personalized PageRank (PPR) triple-hop traversal, managed by a 3-tier adaptive query router with Safety Alignment.

• Results: Produces grounded, citation-backed answers with low hallucination, evaluated on a 60-question benchmark using nDCG@10, grounding overlap, and an LLM-as-judge score. Machine Unlearning for Toxic LLMs Python / Jupyter 2026 GitHub: github.com/Khanhhh239/Machine_UnlearningToxic

• Objective: Research and implement 4 unlearning methods to remove toxic generation from an LLM without full retraining.

• Techniques: Implement four core techniques: Ethos (Task Arithmetic + SVD), NPO + RT (Negative Preference Optimisation), DEPN (Neuron Detection), and RMU + RNA (Representation Misdirection).

• Results: NPO+RT: toxicity score 0.19 0.0007 (−99.6%) Perplexity: 14.78 vs 13.88 baseline. Flow Matching — Generative Models: Theory to Practice Python / Jupyter 2026 GitHub: github.com/Khanhhh239/FLOW_MATCHING

• Objective: Implement Flow Matching (Continuous Normalizing Flows) end-to-end, scaling from 2D toy distributions to conditional CIFAR-10 image generation.

• Techniques: Vanilla + Rectified Flow with EMA on 8-Gaussian / Swiss Roll; evaluated with paper-level metrics

(Wasserstein-2, MMD, Straightness, NFE). Latent FM-DiT pipeline for CIFAR-10 with Classifier-Free Guidance and RK45 ODE solver; top-5 nearest-neighbour audit to rule out memorisation.

• Results: Toy (2D): Rectified Flow straightened trajectories from 0.95 0.99 and path efficiency 0.71 0.99 (kinetic energy 2.31 0.94) near-straight paths, far fewer solver steps to sample. CIFAR-10: Improved pipeline produces class-recognisable images with clear per-class semantic alignment (compared to noisy baseline outputs), staying stable across guidance scales 1.0–4.0.

COMPETITIONS & ACTIVITIES

• AI CiTy 2025 — Track 4: Real-world AI Applications 2026 Text-to-image pedestrian retrieval (including anomalous behavior) on a real-world gallery, trained only on synthetic data — a Sim2Real cross-modal matching problem.

• Bomberland AI Competition — Reinforcement Learning Agent 2026 Design and train an RL Agent for the Bomberland game environment from scratch. AWARDS & ACHIEVEMENTS

Mathematics Olympiad

• 3rd Place — National Olympiad in Mathematics (Grade 11).

• Win a Gold Medal at the Southern Olympiad and a Silver Medal at the Northern Olympiad (Grade 11).

• Admit directly to the AI major (no entrance exam) based on these national achievements.

Contact this candidate