I am Ho Thinh Hung, a Senior Edge AI Engineer working where multimodal models meet physical machines. I optimize Vision-Language-Action and Edge LLM systems, build compact VLA models, and turn research ideas into measurable deployment trade-offs.
model quality × memory × latency × power → useful intelligence at the edge
- Small VLA models for responsive, resource-constrained robotic systems.
- Inference optimization across BF16, FP8, NVFP4, INT8, and INT4/AWQ.
- Edge deployment paths using TensorRT, TensorRT Edge-LLM, ONNX, CUDA, and Jetson.
- Reproducible benchmarks that expose the real memory, latency, throughput, and accuracy trade-offs.
- Learning in public through experiment reports, paper notes, and practical implementation guides.
| Signal | What it represents |
|---|---|
| 100K+ model downloads | Quantized models adopted through a company Hugging Face organization. |
| TensorRT Edge-LLM investigations | Isolated numerical and export failures with controlled layer-by-layer experiments. |
| NVIDIA maintainer confirmation | Findings and fix direction acknowledged in the upstream repository. |
| Production AI systems | Experience spanning GPU inference, real-time multimodal pipelines, and robotics. |
Read the full experiment: TensorRT Edge-LLM — four fixes from controlled experiments
Upstream evidence: issue #151 · issue #105
| Project | Focus |
|---|---|
| Model Quantization Recipes | Practical ModelOpt recipes and benchmark comparisons for BF16, FP8, NVFP4, INT8 SmoothQuant, and INT4 AWQ. |
| TensorRT Edge-LLM | Working fork used to reproduce, isolate, and validate edge LLM/VLM inference failures. |
| Research Note Agent | A workflow for reading papers and publishing engineering-focused, implementation-ready notes. |
| Portfolio & Learning in Public | Interactive benchmarks, project stories, and long-form notes about VLA and model optimization. |
More about the systems I work on
- VLA & multimodal: vision encoders, language backbones, action heads, policy inference, asynchronous execution.
- Optimization: quantization, mixed precision, calibration, KV-cache and activation memory, kernel/runtime profiling.
- Serving: vLLM, Triton Inference Server, streaming APIs, multi-GPU inference, latency and throughput analysis.
- Perception: DeepStream, YOLO, tracking, face recognition, OCR, and real-time video analytics.
- Systems: CUDA, WebRTC, Kafka, Redis, Docker, REST, WebSocket, and production observability.
Small models. Big machines.
I care about the numbers between a paper result and a reliable deployed system.
I care about the numbers between a paper result and a reliable deployed system.