I work on AI projects where a working demo is only the first step. I care about baselines, failure cases, and honest evaluation—then I turn the useful part into something people can run.
Exploring graph learning · recommendation systems · model evaluation · practical AI tools
Open to algorithm, data science, and applied AI internships
Build something real. Measure it honestly. Ship the decision. RankGuard evaluates what should ship, Trend2Video Pro builds a reviewable AI workflow, and WorldCupROI turns mixed evidence into a decision prototype. Each project follows the same standard: reproducible input → inspectable pipeline → measured output → explicit limits.
01 / Evaluate · RankGuard
MovieLens interactions → four recommenders → cohort diagnostics → release gate
An evidence-first recommendation workbench built with PyTorch, FastAPI, React, and TypeScript. It
compares Popularity, Item-kNN, BPR-MF, and LightGCN under one chronological full-catalog protocol.
Item-kNN improved overall NDCG@10 by 57.7% over popularity, but regressed for deep-history users—so
the release gate returned REVIEW instead of hiding the failure behind an average.
Open the live review · Inspect the evaluation · Read the model card
Evidence boundary: offline MovieLens ranking metrics do not establish online CTR, retention, or business lift.
02 / Create · Trend2Video Pro
Trend or URL → scoring and strategy → script and storyboard → MP4 publish package
A local agent workflow that produces a video, subtitles, thumbnail, metadata, and quality report for manual review. The project includes a CLI, FastAPI endpoints, a Streamlit execution console, and SQLite trend memory.
Watch the demo · Inspect the pipeline · Run it locally
Evidence boundary: the no-key demo uses mock LLM responses; its benchmark does not establish real-world content performance.
03 / Decide · WorldCupROI
Historical matches + source text + commercial proxies → baselines and intervals → scenario dashboard
A reproducible sponsorship analytics prototype that joins match context and attention signals, evaluates match and ROI-proxy baselines, and shows uncertainty and scenario comparisons in a dashboard.
Open the dashboard · Read the case study · Check the data card
Evidence boundary: sponsor spend, conversion, and ROI are constructed demo variables. Proxy-model scores do not validate real commercial return or campaign lift.