OPEN TO PH.D. / AI RESEARCH ROLES (2026)

AI Researcher & Systems Engineer

Vo Thanh
Nguyen

Researching generative world models, latent video forecasting, and foundation models for embodied perception. Engineering high-throughput pipelines, multi-GPU training systems, and real-time edge autonomous driving stacks.

Conducted under the mentorship of Dr. Ha Duc Bui (HCMUTE, Ph.D. from NUS).

Download CV (PDF) β†— Contact Me
Google Scholar β†— OpenReview β†— GitHub β†— LinkedIn β†—
πŸ₯‡ 1st Worldwide 1st Place NVIDIA AI City Challenge (Track 2: Scene Understanding)
🏎️ Global Finalist Top 9 Bosch Future Mobility Finals (Romania, Best in Asia)
πŸ“„ Peer-Reviewed 4 Papers ECCV-W 2026, IEEE/CVF WACV 2027 & IEEE GTSD
πŸ’‘ Software Lead 150M VND Bosch Team Sponsorship & Talent Scholarships

Research

Recent Work & Publications

01
Surveillance Retrieval Β· Metric Learning Β· Calibration Β· 2026–2027 β˜… IEEE/CVF WACV-W 2027 Β· Accepted

LENS: Likelihood Evaluation and Neighbor Smoothing for Attribute-Based Person Retrieval on Unseen Cameras

Nguyen Dang Duy Tan, Vo Thanh Nguyen, Linh T. T. Tran, Huynh Nguyen Phat

Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 2027 Β· arXiv:2610.33641

Formulates attribute-based person retrieval on unseen surveillance cameras through likelihood scoring and zero-cost transductive neighbor smoothing:

  • Exact-Query Log-Likelihood Scoring: Replaces heuristic L1 distance ranking with calibrated probabilistic log-likelihood evaluation over multi-teacher attribute predictions.
  • Group Cross-Entropy Training: Models mutually exclusive semantic attribute groups to suppress conflicting predictions under cross-domain camera shifts.
  • Zero-Cost Transductive kNN Smoothing: Reuses forward-pass DINOv3 patch tokens to smooth gallery logits across nearest neighbors without extra backbone inference.
LENS retrieval pipeline showing likelihood evaluation, group cross-entropy, and test-time gallery neighbor smoothing πŸ” Click to expand (HD)
Figure 1 (Retrieval Pipeline): (a) Multi-teacher attribute models with group CE. (b) Transductive kNN smoothing reuses DINOv3 patch tokens from the same pass.
  • Exact-query log-likelihood scoring over traditional L1 metric formulation
  • Zero-cost test-time transductive kNN logit smoothing via DINOv3 patch tokens
  • Paper accepted at IEEE/CVF WACV-W 2027 (68.17 mADM on UPAR 2027)
arXiv β†— Accepted PDF
02
World Models Β· Video Diffusion Β· Generative AI Β· 2026 β˜… ECCV-W 2026 Β· Poster πŸ₯‰ Rank 3 Worldwide (Score: 75.13)

JEPA Guided Diffusion: Predictive Vision-Language Conditioning for Generative Traffic Forecasting

Trinh Tra Giang Nguyen*, Thanh Nguyen Vo*, Nguyen Hoai Thuong Bui*, Ha Duc Bui†

Proceedings of the European Conference on Computer Vision Workshops (ECCVW), 2026

Decouples traffic-world dynamics understanding from pixel-level video synthesis by bridging frozen V-JEPA 2.1 with frozen Cosmos-Predict2.5:

  • Latent World Representation: Extracts spatiotemporal traffic dynamics from observed video clips using frozen self-supervised V-JEPA 2.1 encoders.
  • Lightweight Predictor Bridge: Trains a QFormer and Llama-3.2-1B alignment module with Hungarian set-matching loss to predict future cross-attention conditioning tokens.
  • Generative Video Continuation: Guides frozen Cosmos-Predict2.5 diffusion synthesis without running heavy 7B VLM reasoning at inference (Rank 3 Worldwide, ECCV-W 2026).
Inference overview of the JEPA-guided diffusion model for generative traffic forecasting πŸ” Click to expand (HD)
Figure 1 (Inference Pipeline): The trained predictor bridge replaces Cosmos-Reason1, directly supplying predicted cross-attention context tokens from frozen V-JEPA 2.1 to the frozen Cosmos-Predict2.5 video diffusion generator.
  • Decoupled forecasting framework combining frozen V-JEPA & Cosmos-Predict2.5
  • Lightweight predictor bridge trained with Hungarian set-matching loss
  • Paper accepted at ECCV-W 2026
arXiv β†— Accepted PDF Code
03
World Models Β· Video-LLM Β· Scene Understanding Β· 2026 β˜… ECCV-W 2026 Β· Oral Presentation πŸ† 1st Worldwide Benchmark

Sim-to-Real Traffic Scene Understanding by Decoupling Semantics from Caption Generation with V-JEPA

Nguyen Hoai Thuong Bui*, Thanh Nguyen Vo*, Trinh Tra Giang Nguyen*, Ha Duc Bui†

Proceedings of the European Conference on Computer Vision Workshops (ECCVW), 2026

Bridges sim-to-real traffic scene understanding by decoupling semantic fact extraction from vision-language caption generation:

  • V-JEPA 2.1 Fact Extraction: Encodes multi-view traffic event clips with frozen V-JEPA 2.1 and scores structured VQA candidates via InfoNCE alignment.
  • Training-Free Relational Refinement: Enforces domain-consistent Answer Priors, Question Relations, and Temporal Phase Transitions across predicted attributes.
  • Grounded Caption Generation: Feeds verified semantic facts and keyframe sequences into LoRA-adapted Qwen3-VL-8B (1st Place Worldwide, ECCV-W 2026 Oral).
V-JEPA 2.1 video representation and question token embedding for answer prediction πŸ” Click to expand (HD)
Figure 1 (Answer Prediction): Selected camera event clips are encoded by frozen V-JEPA 2.1 and mapped into latent visual tokens, fused with Llama-3.2 question embeddings for InfoNCE answer scoring.
  • Decoupled semantic fact extraction from caption generation via V-JEPA 2.1
  • Training-free relational & temporal refinement guiding Qwen3-VL-8B
  • Paper accepted at ECCV-W 2026 (Oral Presentation)
arXiv β†— Accepted PDF Code
04
Autonomous Driving Β· Trajectory Planning Β· Real-Time Systems Β· 2026 β˜… IEEE GTSD 2026

An Integrated Perception-to-Planning Framework for Autonomous Driving using YOLOv11 and Frenet

Thanh Nguyen Vo*, Trinh Tra Giang Nguyen*, Lu Anh Khoi Le, Ha Duc Bui†

2026 8th International Conference on Green Technology and Sustainable Development (GTSD), IEEE, 2026

Real-time multi-task perception-to-planning architecture deployed on NVIDIA Jetson Orin Nano for autonomous urban driving:

  • Shared YOLOv11 Backbone: Jointly executes multi-scale traffic object detection and drivable corridor feature extraction in a single forward pass (35+ FPS).
  • Attention Heatmap Prior: Trains a full-resolution spatial attention branch with Dice and Focal loss to anchor trajectory prediction along sharp road curvature.
  • Command-Gated Frenet/BΓ©zier Planning: Modulates waypoint regression with high-level navigation commands to synthesize smooth, kinematically feasible trajectories.
Overview of the proposed dual-branch architecture with shared YOLOv11 backbone, driving branch, and perception branch πŸ” Click to expand (HD)
Figure 1 (Network Architecture): Shared YOLOv11 stem branching into an independent Driving branch (CSM command gating, heatmap prediction, BΓ©zier trajectory head) and Perception branch (multi-scale YOLO detection).
  • Published in IEEE GTSD 2026 Proceedings
  • Command-Gated trajectory planning with Dice loss attention prior
  • Deployed on NVIDIA Jetson Orin Nano with 35+ FPS multi-task execution
IEEE Xplore β†— Accepted PDF

Featured Projects

Engineering Projects

Beyond academic publications, I architect production-grade machine learning platforms, real-time embedded autonomous vehicle stacks, and enterprise recommendation pipelines.

Phoenix ML Platform Architecture
Nov 2024 – Jul 2025 Β· Distributed MLOps & DDD β˜… Production MLOps Platform

Phoenix ML Platform β€” Real-Time Inference & Self-Healing MLOps

Four-layer Domain-Driven Design MLOps platform across 6 bounded contexts and 14 containerized services:

  • Multi-Engine Serving: 5 ML domains served via ONNX, TensorRT, Triton, and BentoML (0.72 ms p99).
  • Feature & Stream Plane: Redis 7 online store (0.21 ms p50), Parquet/DVC, and Kafka + PyFlink.
  • Self-Healing Loop: KS/PSI/χ²/W₁ drift guards auto-trigger Airflow + Optuna retraining in 17.2s.
4-Layer DDD Β· 6 Contexts 5-Domain Model Zoo 0.72 ms P99 Latency 628 Automated Tests
View Interactive Showcase & Live UI β†’
Oct 2024 – May 2026 Β· Autonomous Driving β˜… Rank 1 Vietnam Β· Top 9 Global

Bosch Future Mobility Challenge (BFMC 2025 & 2026)

  • Zero-Copy Runtime: Jetson Orin Nano, RealSense D435i RGB-D, STM32 MCU, and POSIX SHM rings.
  • Multi-Task Perception: 21.8 ms TensorRT NeuroPilot network, FDAT loss, and Joseph-form EKF.
  • 23-State FSM & Control: 9 maneuver handlers with Clothoid G1 + PD-Pure Pursuit at 30 Hz.
21.8 ms TensorRT 23-State FSM 150M VND Grant
View Track Demo Video & Architecture β†’
Jan 2026 – Mar 2026 Β· Large-Scale RecSys β˜… Top 5 Finalist Datathon

Insight Hunter: Chợ Tα»‘t Real Estate RecSys

  • Streaming Forensics: Out-of-core Polars processing 161.7M clickstream logs across 2.09M listings.
  • Hybrid Retrieval: ContactALS (128D) + Vietnamese SBERT + Haversine cold-start BallTree KNN.
  • LambdaRank & MMR: +33.5% NDCG@10 lift, 11.5Γ— catalog coverage, and 38.2 ms TreeSHAP serving.
+33.5% NDCG@10 161.7M Events 11.5Γ— Coverage Lift
View Interactive Showcase & Architecture β†’
Aug 2025 – Feb 2026 Β· Agentic AI & GraphRAG β˜… 6 Specialist Agents Β· 241 Endpoints

KnowledgeForge Enterprise Agentic OS

  • Parallel Topological DAG: 4-lane complexity router & 6 specialist agents with bwrap sandbox.
  • Hybrid GraphRAG: Qdrant dense + BM25 + LightRAG/Neo4j ontology with Cross-Encoder reranking.
  • 4-Tier Memory & TrustLayer: Redis/Qdrant/Neo4j/750+ skills + Presidio PII & 3-Tier RBAC.
241 FastAPI Endpoints 88.4% Cache Hit 1,593 Automated Tests
View Interactive Showcase & Live UI β†’
Mar 2024 – Nov 2024 Β· Zero-Copy POSIX IPC β˜… 49.93 FPS Β· 0.12 ms IPC

Real-Time Driver Distraction Monitoring

  • Zero-Copy IPC: 5 lock-free /dev/shm ring buffers running at 49.93 FPS (TensorRT FP16).
  • 6-Signal Fusion: 478-point FaceMesh EAR/PERCLOS, 6-DoF SolvePnP pose, and YOLOv8-cls behavior.
  • Hysteresis Gate: 3-state temporal alert FSM eliminating single-frame blink false alarms.
49.93 FPS Real-Time 0.12 ms /dev/shm 99.51% Top-1 Acc
View Interactive Showcase & Live HUD β†’
Sep 2025 – Jan 2026 Β· Financial Time-Series β˜… 636K VND Best MAE

Revenue & COGS Forecasting

  • Zero-Leakage Forensics: Reconciles 12 ERP tables with 6-harmonic Fourier seasonality profiles.
  • 3-Tier Ensemble: Blends Q-Specialist LightGBM + L2 Ridge (0.70), GBDT (0.15), and Chronos-2 (0.15).
  • Regime Calibration: Odd/Even-year Walsh COGS calibration cutting MAE by 42.2% to 636K VND.
70/15/15 Ensemble RΒ² = 0.880 Β· 636K MAE 548-Day Horizon
View Interactive Showcase & Architecture β†’
Jun 2026 – Oct 2026 Β· Multimodal Vision β˜… 0.8001 UPAR Challenge Avg

UniQuery: Cross-Domain Pedestrian Attribute Recognition

  • Dual Foundation Encoders: Co-adapts DINOv3 (ViT-H+/16) spatial tokens and CLIP (ViT-L/14) semantic patches.
  • Dynamic Token Routing: Pre-LN cross-attention over 228 tokens routes color to CLIP and clothing/geometry to DINOv3.
  • Cross-Camera Generalization: Outperforms static ensembles across 40 attributes on PA100k, PETA, and Market-1501.
DINOv3 + CLIP 228 Patch Tokens 0.8001 Challenge Avg
View Interactive Showcase & Architecture β†’
Explore All 8 Engineering Projects β†’

Experience

Work & Education

AI Engineer & Researcher

University Research Lab Β· HCMUTE

World Models Autonomous Systems Edge AI
  • Software lead for Bosch Future Mobility Challenge (BFMC 2025 Global Top 9 in Romania & BFMC 2026 Rank 1 Nationally); awarded 150M VND Bosch Research Sponsorship.
  • Researched non-generative world models (V-JEPA 2.1) and diffusion trajectory synthesis for end-to-end autonomous driving and cross-domain surveillance.
  • Authored and co-authored 4 peer-reviewed papers accepted across ECCV-W 2026 (Oral & Poster), IEEE/CVF WACV-W 2027, and IEEE GTSD 2026.

AI Engineer

AIoT & Semiconductor β†—

Industrial Vision POSIX IPC TensorRT
  • Architected deep learning visual inspection pipelines for BACKER printed circuit boards, achieving 99.2% defect classification accuracy.
  • Engineered a multi-process zero-copy POSIX shared-memory runtime across TensorRT, OpenVINO, and ONNX backends, achieving a 3.2x CPU throughput speedup.
  • Delivered continuous real-time defect triage deployed on industrial edge hardware.

B.Eng. Robotics & Artificial Intelligence

Ho Chi Minh City University of Technology and Education (HCMUTE)

GPA: Top 1% Talent Scholar Expected 2026
  • Third Prize ICPC National Contest (2024); Second Prize Vietnam National Olympiad in Informatics (OLP 2023, 2022); Top 5 Finalist VinUniversity Datathon 2026.
  • Recipient of the HCMUTE Talent Scholarship and Certificate of Merit from the University Rector for scientific research and competitive programming achievements.

Specialized Informatics Student

Le Quy Don High School for the Gifted

Informatics Major Algorithms & DS
  • Three-year specialized training in data structures, graph algorithms, dynamic programming, and computational complexity in C++ and Python.