21 projects: 6 production systems at Nam Viet Media (code is proprietary), plus research and coursework, 12 of them with public code.
21 projects shown
Multi-stage news recommender in production: DSSM two-tower retrieval with a BERT-style item tower, NRMS/LSTUR ranking, LLM summaries, recency-aware re-ranking.
Spotify-style music recommender that tags songs from their waveforms (MuQ, MERT, AST, PANNs, CLAP) and models listening sessions with SASRec and BERT4Rec.
Data platform under both recommenders: GA4 events into S3, Bronze/Silver/Gold on Spark + Iceberg, Airflow 3 and Prefect on Kubernetes, on-prem Linux servers.
A project in active development. Details will be published at launch.
B.S. thesis. Removes specific users' data from a trained recommender without full retraining: matches retraining accuracy (~69.3%) with ~5× less compute.
About 10 parallel n8n workflows cross-check API results against crawled pages, score each source on speed and accuracy, and alert via Google Chat.
SerpAPI keywords → GPT outline → editor approval → article in the same thread to save tokens → two-LLM review → CMS draft.
Scrapers for JavaScript-heavy sites with Playwright and Selenium, plus live Chrome sessions driven over a remote debugging port to capture network data.
Full-stack movie recommender for CSE 532: ~1,000 movies from the TMDB API in a 17-table PostgreSQL 16 schema, served as 3 Docker Compose services.
Symptom tracking for Myasthenia Gravis patients: MG-ADL assessments, D3.js trend charts, PDF summaries, and emergency alerts to doctors.
MIND multi-interest retrieval with capsule-style dynamic routing and label-aware attention: 83.92% validation accuracy, HR@50 of 0.104 on MovieLens 1M.
Dual-tower DSSM for user-item matching on MovieLens: 81.74% test AUC and 74.56% test accuracy.
Matrix factorization and SVD on MovieLens 1M with hand-written SGD and mini-batch optimizers: SVD test RMSE 0.2258 vs. 0.2275 for MF.
Transformer encoder-decoder built from scratch, without Hugging Face, trained on Multi30k with BLEU evaluation against a seq2seq baseline.
SimCLR with NT-Xent loss on ResNet-18/50 over CIFAR-10: ~96.5% top-5 test accuracy using only 10% of labels.
Two-stage pipeline: YOLOv8 detects people, the box expands to capture context, a fine-tuned ResNet-50 classifies the activity. 96% test accuracy on Stanford40.
~2,640 country-year observations analyzed with statistical, ML, clustering, and forecasting methods; Ridge Regression reached 0.989 test R².
Python app for a Big Data seminar that scrapes social and YouTube content and analyzes it with Gemini and GPT through a web interface.
Web app with a Python back end that uses Hugging Face language models to correct and extend user-written text.
Data science capstone: API and web-scraping collection, SQL EDA, Folium maps, and a Plotly Dash app for launch-success prediction.
Entry for the Kaggle competition Predict Student Performance from Game Play.