An independent Metal renderer for Wallpaper Engine scenes — not affiliated with Wallpaper Engine. Also plays video and sandboxed web-page wallpapers, with each connected display running its own source. Menu-bar first, with playlists, scheduled rotation, and auto-pause on full-screen, game windows, occlusion, and battery. No accounts, no telemetry. Lite and Pro are split at compile time (#if LITE_BUILD) rather than by hiding UI.

Deterministic profiling and validation come first; LLM workflows extend those facts under constraints. Every claim in a generated report is backed by a claim ledger and checked by hard validators, and open-ended Python analysis runs isolated in a sandbox. Pydantic-typed artifacts end to end, DuckDB for multi-table join discovery and validation.

Click anywhere on a map of New York and get a data-driven dossier on safety, transit, amenities, and building conditions — scored against the city, trended over time, narrated by an on-device LLM, with no cloud calls. One vLLM stack co-tenants Nemotron-30B (NVFP4) and Qwen3-Embedding-4B on a single DGX Spark GB10 alongside a ~30 GiB cuDF/cuVS dataset cache, while cuML DBSCAN clusters incident hotspots across 17 NYC open-data sources.

Two training modules for medical students. An AI patient you interview from first question to final diagnosis, scored on a 6-dimension rubric with live coverage tracking. And camera-based CPR practice — MediaPipe pose tracking at 33 landmarks / 30 fps, compression detection from wrist-motion peaks, scored on 7 dimensions against the full 30:2 BLS protocol.

I owned hybrid retrieval — BGE-M3 dense + sparse over Milvus, SQLite FTS, and web search — and a 5-metric evaluation harness covering tool recall/precision, parallel-call correctness, keyword hit rate, and hallucination, which regression-tests every change.

AI Engineering Intern — China-ASEAN Information Harbor  ·  Jan – Mar 2026

Shipped production Dify workflows extracting structured fields from customs and certificate documents (image + PDF), and built an internal RAG assistant — intent recognition, multimodal extraction, tool routing — over a local corpus with zero external data egress.