I work across the full AI product lifecycle—from data preparation and model training to backend inference APIs, edge deployment, and user-facing applications. My current interests include computer vision, speech and audio intelligence, large language models, retrieval-augmented generation, reinforcement learning, recommender systems, and applied machine learning.

I studied Artificial Intelligence at FPT University, Can Tho Campus (2022–2026).

AI and data
Python · PyTorch · TensorFlow/Keras · scikit-learn · OpenCV · MediaPipe · FAISS · LangChain · Librosa · pandas · NumPy

Application development
FastAPI · Flask · REST APIs · WebSocket · React · TypeScript · Vite · Tailwind CSS · Supabase · SQL

Engineering workflow
NVIDIA Jetson · Ubuntu · Git · GitHub · Jupyter · Model evaluation · Data visualization · API integration

A voice-and-vision admissions assistant I deployed on a Jetson Orin Nano Super. The final system combines a Vietnamese RAG pipeline, streaming speech interaction, CSI-camera face tracking, optional serial hardware control, a FastAPI WebSocket service, and an animated React kiosk interface.

Jetson RAG FAISS Computer Vision Speech AI FastAPI React

A team-built exam-support platform combining a production RAG pipeline, temporal document handling, evidence-aware responses, OCR ingestion, authentication, administration, and a responsive web interface. As Team Lead and Primary Developer, I implemented most of the system, designed the hybrid RAG architecture and its integration, and built the evaluation suite.

RAG LangChain FAISS BM25 FastAPI React PostgreSQL

A research companion to FPT Exam Assistant with reproducible ablations for context deduplication, temporal freshness, automatic document-family discovery, OCR-aware ingestion, and official ParentDocument baselines. It publishes curated aggregate results and clustered-bootstrap analyses while excluding private institutional data.

Information Retrieval RAG Evaluation Temporal Retrieval Python Jupyter

An end-to-end emotion-aware music recommendation application. I trained and integrated facial and speech emotion-recognition models, implemented the Flask inference and recommendation logic, and connected the AI backend to a React/TypeScript interface with Supabase authentication and history.

TensorFlow Computer Vision Speech Emotion Recognition Flask React TypeScript

An intelligent Caro / Gomoku ($7 \times 7$ Connect-4) agent trained via Proximal Policy Optimization (PPO) and Actor-Critic SE-ResNet architectures. It incorporates curriculum self-play with an opponent pool, 8-fold dihedral data augmentation, and an anti-collapse safeguard to achieve a 98.1% win rate against tactical heuristic bots. Includes a real-time desktop GUI for human vs. AI evaluation.

PyTorch Reinforcement Learning PPO Actor-Critic Self-Play Python Tkinter

A privacy-conscious camera attendance prototype with local face enrollment, real-time recognition, and timestamped Excel reporting.

Python MediaPipe OpenCV NumPy openpyxl

An item-based collaborative-filtering system using sparse rating matrices and KNN, with experiments comparing cosine, Pearson, MSD, L1, and L2 approaches using RMSE.

Python scikit-learn Surprise SciPy pandas

An end-to-end traffic-density modeling workflow with exploratory analysis, candidate-model comparison, model export, and a validated Flask inference application.

Python scikit-learn XGBoost Flask pandas

• Building AI features that solve a clear user problem.
• Evaluating models honestly and documenting their limitations.
• Turning notebooks into reproducible applications and APIs.
• Keeping credentials, private datasets, and personal data out of source control.

transtoanvovankiet@gmail.com