Production Infrastructure & Applied AI

Engineering Projects

Production-grade MLOps platforms, real-time autonomous driving stacks, large-scale recommendation engines, and zero-copy multiprocessing runtimes designed for low-latency reliability.

⚡ 0.72 ms Inference · 4.8 ms HTTP E2E 🏎️ Jetson Orin Nano · 21.8 ms Multi-Task Perception 🎯 161.7M Clickstream Events · 11.5× Coverage ⚙️ Zero-Copy POSIX Shared-Memory IPC 🎬 Interactive Visual & Video Walkthroughs

Oct 2024 – May 2026 Autonomous Driving · Embedded C++20 & TensorRT ★ 1st Nationally (2026) ★ Global Top 9 (2025, Best in Asia) 150M VND Bosch Grant

Bosch Future Mobility Challenge (BFMC 2025 & 2026)

1:10-scale autonomous driving stack on NVIDIA Jetson Orin Nano uniting zero-copy C++20/Python IPC, 21.8 ms TensorRT multi-task perception, Joseph-form EKF localization, and a 23-state behavior FSM.

21.8 ms TensorRT FP16 Multi-Task

3.8 cm Lateral Tracking RMSE

99.2% 120-Run Track Completion

30 Hz 23-State FSM & STM32 Loop

Sensing & Zero-Copy IPC

→

Streams RGB-D frames and 100 Hz IMU telemetry via lock-free POSIX shared memory and C++20 worker threads.

RealSense D435i POSIX /dev/shm ZeroMQ Bus

NeuroPilot & EKF Fusion

→

Joint 13-class detection, FiLM-routed cubic Bézier waypoints (FDAT loss, -41% error), and Joseph-form EKF pose.

YOLOv11 + BiFPN FDAT Loss Joseph EKF

23-State FSM & Actuation

Orchestrates 9 urban maneuvers with Clothoid G1 smoothing, PD-Pure Pursuit steering, and STM32 watchdog safety.

23-State FSM Clothoid PDPP STM32 Nucleo

Closed-Loop 1:10 Autonomous Track Run with NeuroPilot 10-Waypoint Trajectory and Traffic Sign Detection — **Figure 1 (Closed-Loop 1:10 Track Run — NeuroPilot Waypoint & Sign HUD):** Real-time onboard deployment executing 10-waypoint cubic Bézier trajectory tracking, dynamic steering control (33.0 cm/s), and 3D traffic-sign/stop-line distance estimation.

Jan 2026 – Mar 2026 Recommender Systems · Large-Scale Retrieval 🏆 Top 5 Finalist — VinUniversity Datathon 2026

Insight Hunter: Large-Scale Real Estate Recommendation System

4-stage hybrid retrieval and diversity-aware reranking pipeline trained on 161.7M Chợ Tốt clickstream logs and 2.09M property listings, lifting NDCG@10 by +33.5% and catalog exposure by 11.5×.

161.7M Clickstream Events (12 GB Cap)

+33.5% NDCG@10 Ranking Lift

11.5× Catalog Exposure (42.6%)

38.2 ms p99 TreeSHAP Slate Serving

Streaming & Tri-Modal Recall

→

Out-of-core Polars matrix weighting high-intent leads (6.0×) with ContactALS, VN-SBERT, and Haversine BallTree KNN.

Polars LazyFrame ContactALS 128D VN-SBERT + KNN

Domain Guard & LambdaRank

→

Prunes 91.9% out-of-zone listings via post-2025 ward boundaries and scores Top-500 candidates across 42 features.

Ward Boundary Gate Legal Title Filter LightGBM LambdaRank

MMR Fairness & Explainability

Cuts seller Gini monopoly by 38.4% across 415K sellers and serves Top-10 slates with real-time TreeSHAP attributions.

MMR (λ=0.7) 415K Seller Reach TreeSHAP API

Insight Hunter 4-Stage Hybrid Retrieval and Reranking Architecture — **Figure 1 (4-Stage Pipeline Architecture):** Out-of-core Polars feature extraction, hybrid ContactALS + VN-SBERT + Haversine retrieval, domain rule filtering, and LightGBM LambdaRank + MMR reranking.

Aug 2025 – Feb 2026 Agentic AI · Multi-Agent DAG & Hybrid GraphRAG 6 Specialist Agents · 241 Endpoints · 1,593 Tests

KnowledgeForge: Enterprise Agentic OS & Hybrid GraphRAG Platform

Enterprise Agentic AI Operating System pairing a Next.js 15 / React 19 workspace with a 241-endpoint FastAPI control plane, 6-agent topological DAG orchestration, tri-modal GraphRAG retrieval, and 4-tier cognitive memory.

241 APIs FastAPI Control Plane & RBAC

6 Agents Parallel Topological DAG Waves

88.4% 5-Layer Prefix Cache Hit Rate

1,593 Tests RAGAS Claim Faithfulness ≥ 0.85

TrustLayer & Hybrid GraphRAG

→

Sanitizes 14-channel inputs via Presidio PII masking and fuses Qdrant dense, BM25 lexical, and Neo4j/LightRAG hops.

Presidio PII + RBAC Qdrant + BM25 + Neo4j MinHash + Cross-Encoder

6-Agent Topological DAG

→

Routes tasks across 4 complexity lanes into parallel specialist waves isolated inside zero-egress Bubblewrap sandboxes.

LangGraph DAG Bubblewrap Sandbox 9Router Multi-LLM

4-Tier Memory & Claim Gate

Persists T1–T4 cognitive memory (Redis, Qdrant, Neo4j, 750+ skills) with DeBERTa NLI atomic citation verification.

T1–T4 Memory DeBERTa Claim Gate OpenTelemetry UI

KnowledgeForge Enterprise Agentic OS Architecture — **Figure 1 (System Architecture):** End-to-end architecture connecting Next.js 15 / FastAPI TrustLayer ingress, LangGraph topological multi-agent orchestration, hybrid GraphRAG retrieval, 4-tier cognitive memory, and OpenTelemetry observability.

Mar 2024 – Nov 2024 Edge Computer Vision · Zero-Copy POSIX IPC 49.93 FPS Real-Time · 0.12 ms Shared-Memory IPC

Real-Time Driver Distraction & Fatigue Monitoring System

Zero-copy multiprocessing driver safety runtime bypassing Python's GIL via POSIX shared-memory rings to fuse 478-point FaceMesh biometrics and TensorRT YOLOv8-cls behavior recognition at 49.93 FPS.

49.93 FPS End-to-End Pipeline Throughput

0.12 ms Lock-Free /dev/shm Ring IPC

99.51% 10-Class Distraction Top-1 Acc

-68% False-Positive Blink Reduction

Zero-Copy POSIX SHM Bus

→

Decouples camera DMA capture, CPU geometric workers, and GPU inference across 5 lock-free shared-memory rings.

5 /dev/shm Rings Zero Serialization True GIL Bypass

Dual CPU/GPU Biometric Workers

→

Runs 478-pt FaceMesh EAR/MAR/PERCLOS + 6-DoF SolvePnP head pose alongside TensorRT FP16 YOLOv8n-cls.

MediaPipe 478-Pt cv2.solvePnP 6-DoF TensorRT / OpenVINO

6-Signal Hysteresis & HUD

Suppresses transient blink spikes via a 3-state temporal FSM (1.0s Warning / 1.5s Alert) and 50 FPS OpenGL HUD.

Blink Gate (<300ms) 3-State Alert FSM PySide6 OpenGL HUD

Zero-Copy POSIX Shared-Memory Multi-Process Architecture — **Figure 1 (Zero-Copy Multiprocessing IPC):** Four-stage pipeline linking lock-free `/dev/shm` ring buffers, parallel MediaPipe & TensorRT workers, 6-signal fusion, and hysteresis alerting.

Sep 2025 – Jan 2026 Financial Time-Series · 3-Tier Ensemble ML 636K VND Best MAE (-42.2% Error)

Enterprise Revenue & COGS Multi-Horizon Forecasting System

Zero-leakage financial forecasting pipeline reconciling 12 relational ERP tables with 6-harmonic Fourier seasonality, a 70/15/15 3-tier ensemble, and odd/even-year COGS regime calibration over a 548-day horizon.

636K VND Expanding-Window MAE (-42.2%)

548 Days P10 / P50 / P90 Forecast Horizon

70/15/15 3-Tier Ensemble Blend Weights

-75% Err Odd/Even Walsh COGS Calibration

ERP Schema & Fourier Profile

→

Reconciles 12 ERP tables into a daily ledger and decomposes annual baseline level from 6-harmonic Fourier waves.

12 ERP Tables 6-Harmonic Fourier Zero-Leakage Lags

3-Tier Ensemble (70/15/15)

→

Blends Quarter-Specialist LightGBM + L2 Ridge (0.70), Era-Weighted GBDT/Huber (0.15), and Chronos-2 120M (0.15).

Q-LightGBM + Ridge Huber + GBDT Chronos-2 (120M)

Walsh Fix & Conformal SLA

Corrects alternating fiscal-year COGS regime shifts via Walsh basis functions and exports 548-day conformal bands.

Walsh COGS Fix R² = 0.880 548-Day Conformal

Revenue and COGS Multiplicative Seasonality and 3-Tier Ensemble Architecture — **Figure 1 (Forecasting Pipeline Architecture):** Six-stage pipeline from 12 relational ERP tables and multiplicative Fourier seasonality to the 70/15/15 3-Tier Ensemble and Walsh COGS calibration.

Nov 2024 – Jul 2025 Distributed MLOps · 4-Layer DDD & Multi-Engine Serving 0.72 ms Compute p99 · 14 Services · 628 Tests

Phoenix ML: Self-Healing Enterprise MLOps & Multi-Engine Serving Platform

4-layer Domain-Driven Design MLOps platform orchestrating 14 containerized services across 5 ML domains with 0.72 ms ONNX/TensorRT serving, dual Redis/Parquet feature stores, and autonomous Airflow + Optuna self-healing.

0.72 ms p99 Compute Latency (ONNX/TRT)

14 Services 4-Layer DDD · 6 Bounded Contexts

1,284 req/s 4 Traffic Routers + Circuit Breaker

17.2 s Airflow + Optuna Auto-Heal Cycle

Streaming & Dual Feature Store

→

Ingests Kafka events via PyFlink into Feast Redis 7 online store (0.21 ms p50) and Parquet/DVC offline storage.

Kafka + PyFlink Redis 7 (0.21ms) Parquet + DVC

CQRS & 4-Router Inference

→

Routes requests via Single, A/B (90/10), Canary (80/20), and Shadow strategies across 4 runtimes with circuit isolation.

CQRS Bus ONNX + TensorRT Triton + BentoML

Drift Guards & Self-Healing

Detects KS/PSI/χ²/W₁ drift to trigger 6 Airflow retraining DAGs, Optuna HPO, and MLflow Champion hot-swaps.

Evidently 4-Test Airflow + Optuna MLflow + Grafana

Phoenix ML 4-Layer DDD and Self-Healing MLOps Architecture — **Figure 1 (Platform Architecture):** End-to-end enterprise MLOps architecture integrating Kafka/Flink streaming, Feast Redis/Parquet feature stores, multi-engine serving, and Airflow + Optuna self-healing.

Jun 2026 – Oct 2026 Multimodal Vision · DINOv3 + CLIP Token Routing 0.8001 UPAR Challenge Avg · 228 Multimodal Tokens

UniQuery: Unifying Vision Models with a Joint Query Head for Cross-Domain Pedestrian Attribute Recognition

Dual-backbone pedestrian attribute recognition system co-adapting DINOv3 (ViT-H+/16) and CLIP (ViT-L/14) with an attribute-specific cross-attention query head over 228 multimodal patch tokens across unseen surveillance camera domains.

0.8001 UPAR 2027 Hidden Test Avg

228 Tokens 99 DINOv3 + 129 CLIP Patches

40 Queries Attribute-Specific Routing

3 Datasets Zero-Shot Cross-Camera Transfer

Dual Foundation Encoders

→

Extracts 99 spatial patch tokens from DINOv3 ViT-H+/16 (24 fine-tuned blocks) and 129 semantic tokens from CLIP ViT-L/14.

DINOv3 ViT-H+/16 CLIP ViT-L/14 228×1024 Tokens

Dynamic Cross-Attention Routing

→

Routes 40 learnable attribute queries via Pre-LN cross-attention: color attributes specialize to CLIP and structural/clothing to DINOv3.

Pre-LN Cross-Attn 40 Attribute Queries Token Specialization

Tri-Head Logit Fusion

Averages logits across two single-encoder query heads and the joint multimodal query head for cross-domain generalization.

3-Head Logit Avg PA100k / PETA / Market 0.8001 Challenge Avg

UniQuery dual foundation encoder architecture with joint and single-encoder query heads — **Figure 1 (System Architecture):** Joint fine-tuning of DINOv3 (ViT-H+/16) and CLIP (ViT-L/14) with three query heads: two single-encoder heads and a joint cross-attention query head over 228 tokens.

Apr 2023 – Oct 2023 Multimodal Search · OpenAI CLIP & Faiss Vector Engine <1 ms Faiss Cosine Search · Dual CBIR + TBIR

Natural Landscape Image Search: Multimodal CLIP ViT-B/32 & Faiss Retrieval Engine

Dual-mode Content-Based (CBIR) and Text-Based (TBIR) visual search engine projecting high-resolution landscapes and natural-language queries into a shared 512-D OpenAI CLIP ViT-B/32 hypersphere indexed with Facebook Faiss.

0.8 ms Exact Faiss Inner-Product Search

512-D L2-Normalized CLIP Hypersphere

6 Domains 25+ Automated Pexels Taxonomies

CBIR + TBIR Zero-Shot Image & Text Queries

Pexels Harvester & Dual Ingress

→

Crawls curated HD landscape corpora across 6 biome taxonomies and accepts image uploads or free-text prompts.

Pexels REST API CBIR Image Upload TBIR Text Prompt

CLIP ViT-B/32 & HDF5 Store

→

Projects visual patches and BPE text tokens into unit-normalized 512-D float32 vectors persisted in HDF5.

CLIP ViT-B/32 L2 Unit Norm HDF5 Tensor Store

Faiss IndexFlatIP & Web UI

Executes exact cosine-equivalent inner-product retrieval in <1 ms and renders interactive Top-K galleries.

Faiss IndexFlatIP Dynamic Top-K Flask Web UI

Natural Landscape Image Retrieval CLIP ViT-B/32 and Faiss Architecture — **Figure 1 (System Architecture):** Four-stage pipeline linking the Pexels REST API crawler, OpenAI CLIP ViT-B/32 dual vision-text encoders, HDF5 + Faiss IndexFlatIP storage, and the Flask CBIR/TBIR web gateway.

Core Stack

Technical Competencies & System Stacks

Comprehensive engineering toolset spanning deep learning research, edge deployment, and production systems.

Deep Learning & Vision

PyTorch V-JEPA 2.1 Cosmos-Predict2.5 DiT Qwen3-VL-8B DINOv3 / CLIP LoRA Adaptation YOLOv11 Hugging Face OpenCV

Edge AI, Robotics & Autonomous Systems

TensorRT (FP16/INT8) Intel RealSense D435i RGB-D C++20 (std::jthread) POSIX SharedMemory IPC ZeroMQ XSUB/XPUB Joseph-Form EKF & ByteTrack 23-State Autonomy FSM Clothoid G1 & PDPP Control STM32 NUCLEO Serial

MLOps, Distributed Serving & Data Systems

FastAPI / gRPC / BentoML ONNX / Triton / OpenVINO Apache Airflow & Kafka KRaft MLflow + MinIO Registry Redis 7 Feature Store Prometheus / Grafana / Jaeger Polars LazyFrame & LightGBM LangGraph & Hybrid GraphRAG Docker & Kubernetes HPA