Tri Nguyen Minh

Developed and deployed high-performance microservices architecture handling 15,000+ RPS using Go, FastHTTP, and Redis. Engineered resilient event-driven messaging with Apache Kafka.

Interactive AI conversational assistant trained on Tri's technical background, projects, and architecture specs. Supports streaming responses and zero-trust Tailscale Funnel HTTPS deployment.

Resolved hardware signal contention and pinout conflicts for multi-servo controls, optimizing physical logic limits. Programmed custom ESPHome logic with automated PWM cooling loops via Home Assistant, eliminating thermal throttling.

Managed a self-hosted Ubuntu 22.04 LTS environment acting as a primary node for containerized deployments. Engineered secure tunneling via Tailscale & NGINX Reverse Proxy.

Integrated Google Gemini 1.5 Pro for dynamic IELTS scoring with sub-second streaming latency. Engineered real-time audio-visual synchronization driving a 3D Sensei Avatar (<50ms delta). Applied SuperMemo-2 SRS spaced repetition and offline-first SQLite sync achieving 99.5% reliability.

Cross-platform mobile personal finance management application built with Flutter/Dart. Features automated expense categorization, interactive visual budget analytics charts, monthly spending thresholds, and offline-first local SQLite storage.

Collaborated to architect a GMP-WHO pharmaceutical management system; defined flexible domain models to guarantee scalability. Implemented JWT/RBAC security, managing 15+ complex entities. Dashboard & Mobile App process real-time updates with sub-200ms latency.

Optimized few-shot voice cloning pipeline using PyTorch, successfully synthesizing high-fidelity audio from just 1 minute of raw reference data while reducing inference latency by 35%. Fine-tuned acoustic models and neural vocoders for exact timbre reproduction.

Engineered an end-to-end facial recognition security pipeline, extracting 128-d vector embeddings with sub-200ms latency via OpenCV/dlib. Executed Euclidean distance matching, achieving 98.5% accuracy across 5,000+ benchmark images at 30+ FPS.

Led research team implementing scalable CURE (Clustering Using REpresentatives) algorithms, processing 100,000+ data points with 25% outlier noise density. Engineered optimal representative point partitioning, outperforming K-Means/DBSCAN by 18% in silhouette score accuracy.

Investigated non-Euclidean architectures using Graph Neural Networks (GNN) and architected relational data schemas spanning 1.5M+ nodes. Conducted rigorous peer code reviews and standardized evaluation pipelines, significantly improving experimental reproducibility.