Nguyen Mai Hong Tram
I build multimodal AI systems for real workflows.
I work across OCR, language models, retrieval, visual reasoning, and browser inference—from implementation through evaluation and delivery.
Selected builds
Systems I built, the decisions behind them, and what can be inspected publicly.
Problem
Help teachers review Vietnamese secondary-school mathematics work with evidence-backed draft feedback and class-level error analysis.
What I built
I personally built the OCR flow, LLM API integration, RAG pipeline, and frontend, and handled project management and documentation.
Engineering approach
A teacher-in-the-loop workflow combines multimodal grading, source-grounded retrieval, confidence-based routing, citation checks, PII safeguards, and human review.
Evidence
The public capstone demo shows the product workflow; the system also uses a golden set for evaluation.
Sign-language inference in the browser
TMA Solutions · AI Engineer internship
Problem
Make continuous sign-language recognition usable in a real-time browser setting.
What I built
I researched continuous sign-language recognition and text-to-3D-avatar generation, then built an end-to-end inference path from PyTorch models to the browser.
Engineering approach
Converted model inference to ONNX Runtime Web and integrated it with a browser-facing pipeline.
Evidence
The pipeline was evaluated in a real-time video-chat setting.
Problem
Can explicit object relationships improve surgical visual question answering?
What I built
I carried this research from problem formulation through data preparation, model development, evaluation, and manuscript writing.
Engineering approach
Combined visual representations, language processing, and graph-based scene reasoning in an end-to-end VQA pipeline.
Evidence
Using SSG-VQA, a dataset of about 960K question–answer pairs, the model achieved 85.9% accuracy and 85.8% weighted F1 on a 77,198-question visual-oracle test.
Air Quality Forecasting Pipeline
Forecasting prototype · UEH course group project
View public code ↗Problem
Forecast air quality from OpenAQ measurements and make predictions available through an API and monitoring dashboard.
What I built
I implemented the technical pipeline end to end: OpenAQ ingestion and preprocessing, RNN/LSTM/GRU training, FastAPI inference, a Streamlit dashboard, and Docker packaging.
Engineering approach
Compared recurrent forecasting models and connected model inference to a dashboard through a containerized API.
Evidence
The public repository contains the implementation, setup steps, and model evaluation. The hosted demo is no longer running.
Engineering experience
Applied work across multimodal pipelines, video understanding, and browser inference.
Jul–Aug 2026
AI Research Intern
VinRobotics · AI Platform, VLA Team
Temporally grounded action captioning in egocentric video, with controlled experiments in sampling, grounding, prompting, and model adaptation.
Mar–May 2026
AI Research Engineer Intern
TMA Solutions
Bidirectional sign-language communication research and browser inference using PyTorch and ONNX Runtime Web.
How I work
Tools appear here in the context of the systems they helped deliver.
Applied AI
OCR, LLM APIs, RAG, and multimodal workflows in GradeMind.
Model development
PyTorch, vision-language pipelines, graph reasoning, and controlled evaluation.
Inference and delivery
ONNX Runtime Web, SGLang, FastAPI, PostgreSQL/pgvector, and Docker.
Research foundations
Published work and recognition that inform my engineering practice.
2026
GNN-SurgVQA: Object-Centric Graph Reasoning for Visual Question Answering in Laparoscopic Scene Understanding
International Conference on Data Analytics and Management (ICDAM) · Best Paper · accepted, paper not yet public
Public code ↗2025
Comprehensive Approach to Vietnamese Scene Text Recognition: Challenges, Models, and Framework Development
ICDAM 2025 · Lecture Notes in Networks and Systems, vol. 1601 · Springer · Published
Read paper ↗