I turn papers into experiments, experiments into evidence, and occasionally evidence into useful apps.
Multimedia Retrieval · Multimedia Verification · Computer Vision · Neural Architecture Search · Competitive Programming
CV · Codeforces: Duy_e · Kaggle: meowluvmatcha
Currently retrieving events, verifying media, and making research a little less chaotic.
- Ægis: AI-Enhanced OSINT for Multimedia Verification — ACM Multimedia 2025
Minh-Anh Pham, Anh-Tai Pham-Nguyen, Anh-Duy Le, et al. · An AI-enhanced OSINT system for multimedia verification, forensic analysis, and evidence validation. - MERVIN: A Unified Framework for Multimodal Event Retrieval in Vietnamese News Videos — SOICT 2025
Anh-Tai Pham-Nguyen, Tung-Duong Le-Duc, Anh-Duy Le, Trung-Hieu Truong-Le · A multimodal retrieval framework combining keyframes, transcripts, and video summaries.
- Multimedia retrieval and verification: building systems that search across visual and textual evidence and help verify real-world media.
- Efficient LLM inference — in progress: currently experimenting with serving Qwen3.5-2B using vLLM for the Viettel AI Race 2026.
- Research Navigator: a local-first tool that connects research goals, questions, experiments, decisions, and evidence.
Two recent Kaggle notebooks I'm especially proud of:
- SCRFD+SPOS — NAS and Inference — neural architecture search and inference experiments for SCRFD using SPOS.
- SPOS — Supernet for SCRF Dynamic Block — SPOS supernet experiments with dynamic blocks on the WIDER FACE dataset.
| Project | What it does |
|---|---|
| Research Navigator | A local-first research planning system with evidence tracking, experiment history, a CLI, and MCP support for coding agents. |
| vLLM Colab Inference | An inference optimization workspace for benchmarking TTFT/TPOT, evaluating GPQA accuracy, and shipping a reproducible vLLM setup. |
| NutriScan | A computer vision app that recognizes Vietnamese food with LSNet and turns predictions into nutrition-aware meal suggestions. |
| Speaking Practice App | An AI pronunciation coach using Whisper, text-to-speech, IPA conversion, and similarity scoring. |
| Yolo-ST-GCN | A config-driven skeleton action recognition pipeline with flexible Penn14/COCO18 joint layouts and two-stream ST-GCN training. |
| TensorTonic Solutions | My Python solutions and notes for hands-on tensor and machine learning problems. |
|
Quiz App A small Next.js quiz playground you can try live. Life Management App The app I built after deciding one todo list was clearly not enough. |
|
|
Python · PyTorch · OpenCV · vLLM · React · TypeScript · Node.js · Git
If you're also turning messy ideas into testable systems, we'll probably have something to talk about.