Low-latency, cross-address-space IPC library for ARM that lets Linux processes invoke TrustZone secure services without managing shared memory. Built the transport core (call marshalling, parking/waking over a non-blocking transport, callbacks by reference) behind a socket-like client interface adopted by 5 partner teams.

Rule-driven knowledge retrieval service where AI applies defined rules to surface supporting evidence — a centralized source of truth for company data. Packaged it as reusable Skills, a CLI, and MCP servers, cut load and query latency ~40%, and built an agent pipeline that generates tracking test cases from PRDs at ~80% of manual coverage.

Mission control dashboard and onboard data file system for ELFIN, a NASA space-weather CubeSat. Processed 10M+ telemetry packets during live flight operations.

Parameter-efficient fine-tuning infrastructure using LoRA: trained 7 domain-specific LLM adapters and demonstrated training-free multi-model fusion via subspace orthogonality, with composed models achieving up to 9.1% lower perplexity. Mentored by Prof. Yingnian Wu at UCLA AI Lab.

Full-stack platform that generates topic-specific quizzes from user-uploaded PDFs using LangChain retrieval and ChromaDB embeddings, served via Vertex AI.

Chrome extension that detects difficult words and shows hover definitions, with a customizable word bank. Built in HS to read philosophy faster; became a first-author NLP paper at AIRCC CSIT.