AI Engineering.

I deploy and benchmark LLMs on GPU servers, build coding agents, and integrate AI into web apps. My projects also span multilingual NLP, computer vision, and explainable AI.

From model serving to AI apps

Serving LLMs on GPUs

I build and maintain Biexce’s internal serving platform for coding models. Bifrost routes model requests to vLLM, which runs in Docker on GPU servers. I monitor the services and troubleshoot failures across the stack.

Read the inference case study

INFERENCE REQUEST PATH

Developer tools send requests to a shared API. Bifrost routes each request to the configured model endpoint.

vLLM queues and batches requests for inference. Docker packages the serving runtime for deployment on GPU hosts.

Model weights and cached attention state use GPU memory. That memory budget affects how much context and how many concurrent requests a server can handle.

Benchmark coding models

I benchmark coding models on representative tasks to weigh code quality against response time, throughput, and GPU memory use. A model that performs well on one request can behave differently when several developers use it at once.

See how I evaluate models

BENCHMARK CRITERIA

Code quality
Does the patch satisfy the task and pass the relevant tests?
Latency
Time to first token and time to finish the response
Throughput
Total output tokens per second under concurrent load
GPU memory
Peak VRAM for the same request workload
Stability
Failed or timed-out requests across repeated runs

Build multi-agent workflows

I build multi-agent workflows for planning, implementation, testing, and review. MCP connects the workflow to Unity, Git, and internal docs. A proposed diff and test results give reviewers something concrete to inspect.

See the agent workflow

CODING AGENT WORKFLOW

Use the task and relevant project documentation to define which parts of the project are affected and how to verify the change.

Use Unity and Git context to inspect the affected code and prepare a scoped patch.

Compile the project and run checks tied to the acceptance criteria. A failed check goes back with the steps needed to reproduce it.

Use the diff and test report to check scope, flag regressions, and prepare the change for a developer’s decision.

Integrate AI into web apps

I build web features backed by language models. This request path shows how a feature can validate input on the server and handle usable, incomplete, or failed responses.

WEB APP REQUEST FLOW

The UI gathers input for the feature. The server validates it and builds the model request with the instructions and context the task needs.

Call the model from the backend. Text can stream to the UI, while structured output needs a complete response that passes schema and application checks.

Render the usable result. Give timeouts, incomplete responses, and model refusals distinct states so users know what happened and when retrying makes sense.

Training data quality

THINKWARE VIETNAM · AUG 2024 – SEP 2025

Data Labeling & Review

I moved from labeling road objects to reviewing the team’s annotations before delivery. The projects covered 900,000+ dashcam images from South Korea, Japan, and the United States.

ANNOTATION QA

Apply the project’s labeling rules to vehicles, pedestrians, and traffic signs in dashcam images.

Find missing objects and inconsistent labels before delivery. For example, partially visible objects need the same inclusion rule across reviewers.

Use recurring disagreements and ambiguous examples to clarify the labeling guidelines with the annotation and AI teams.

NLP & computer vision

SOUTH TELECOM · NLPMultilingual Language Detection

I built an XLM-R pipeline for language identification and measured accuracy and inference speed across datasets. The case study shows how I prepare Vietnamese text and why short or mixed-language messages need their own evaluation criteria.

Read the case study
KIENLONGBANK · COMPUTER VISIONSignature detection

I defined annotation guidelines and fine-tuned YOLOv8 to locate signatures in transaction documents. The detector reached 92% precision and 88% recall. I exported it to ONNX and optimized inference with TensorRT, improving inference speed by about 40% in the project benchmark.

HCMUE · XAIExplainable crop disease classification

I was the first author of this study in the TNU Journal of Science and Technology. We added a Convolutional Block Attention Module (CBAM) to VGG16 and measured the overlap between explanation maps and diseased leaf regions using intersection over union (IoU) on selected images.

Read the paper
THESIS · XAIExplainable emotion classification for Vietnamese text

For my undergraduate thesis, I worked on a model that classifies emotions expressed in Vietnamese text. The goal was to make each prediction interpretable by showing what in the text supported the predicted emotion.

COMPUTER VISIONFacial expression recognition for classroom feedback

I built and evaluated a CNN to classify visible facial expressions. The project explored how teachers could use these observations alongside classroom context to reflect on student responses and adjust teaching activities.

Have an AI project?

Model deployment, agent workflows, and AI features for web apps.

Discuss an AI project