Proof, not promises.

Six builds chosen for hiring signal: AI tooling, VLM backend, Vietnamese vision-language research, search, detection, and audio forensics.

01
MCP Browser automation remote gif

MCP Browser

AI Tooling / Browser Automation

AI agents need reliable browser control, but vanilla Playwright-based MCP servers often fail on bot-protected websites due to fingerprinting, Cloudflare, reCAPTCHA, and anti-automation checks. Integrated CloakBrowser stealth Chromium with MCP, humanized interactions, semantic snapshots, multi-session isolation, workflow record/replay, captcha detection, HAR diagnostics, and native video recording in one local server. Architected and implemented a Python FastMCP server exposing 16 browser automation tools for AI…

PythonFastMCPPlaywrightCloakBrowserMCPFFmpeg
02
mp4 demo

PPE Detection System

Computer Vision / Safety Monitoring

Workplace safety checks for helmets, vests, gloves, and boots are often manual, slow, and inconsistent in construction or industrial environments. Extended object detection into per-worker safety reasoning by combining YOLOv8/RT-DETR detections with rule-based logic that checks PPE inside each worker bounding box and flags no_* unsafe classes. Built an end-to-end Streamlit application for video/webcam/uploaded media, model selection, confidence tuning, annotated output export, and SAFE/UNSAFE worker…

YOLOv8RT-DETRPythonOpenCVStreamlit
03
Vietnamese captioning dataset dataset cover

Vietnamese Image Captioning System

Vision-Language / Vietnamese NLP

Vietnamese image captioning is challenging because existing vision-language models have limited Vietnamese support and struggle with diacritics and tone marks. Adapted BLIP for Vietnamese caption generation with dual accented/unaccented pipelines, using PhoBERT for accented Vietnamese and BERT for unaccented text. Prepared training configs, fine-tuned BLIP on UITViC and KTViC, implemented inference with beam search and nucleus sampling, and compared accented versus unaccented approaches. Achieved best overall…

BLIPPhoBERTBERTPyTorchTransformers
04
NutriTrack AWS architecture aws diagram

NutriTrack Backend Microservice

VLM / Backend / Cloud

Manual nutrition tracking is slow and error-prone when users need to estimate food portions, read nutrition labels, or search product data from barcodes. Designed a VLM backend around Qwen3-VL that handles food images, nutrition-label scans, and barcode workflows while reducing token cost through image compression, crop/resize preprocessing, prompt optimization, CSV-style outputs, and barcode/nutrition result caching. Built the backend microservice, third-party API integration layer, deployment configuration, and…

PythonQwen3-VLDockerAWS ECSFly.ioTerraform
05
mp4 demo

Natural Food Vertical Search Engine

Information Retrieval / Semantic Search

General-purpose search lacks precision for domain-specific natural food and biomedical literature, where users need both exact term matching and conceptual retrieval. Combined TF-IDF lexical retrieval, FAISS semantic search, transformer-based spelling correction, lazy model loading, and biomedical SentenceTransformer models in a single Streamlit search interface. Built the indexing, search, evaluation, and UI pipeline for NFCorpus, including inverted index construction, FAISS embedding index, CLI/Streamlit…

PythonTF-IDFFAISSSentenceTransformersStreamlitspaCy
06
Anti-piracy spectrogram spectrogram

Cinema Anti-Piracy Audio Watermark Analyzer

Audio Forensics / Hackathon Prototype

Cinema piracy investigations need a technical way to verify whether leaked online videos contain inaudible high-frequency markers intentionally embedded in theater audio. Turned the anti-piracy idea into an interactive forensic analyzer that extracts audio from videos, visualizes FFT frequency spectra and spectrograms, supports high-frequency inspection, dB scaling, and peak analysis for hidden marker verification. Built the Streamlit UI, FFmpeg video-to-audio pipeline, librosa/NumPy FFT analysis, spectrogram…

PythonStreamlitlibrosaNumPySciPyMatplotlibFFmpeg