Systems I help build
From numerical foundations to model inference and agent applications. These projects combine implementation, controlled experiments and feedback from real users.
A scientific computing and deep learning framework for OCaml. My work spans numerical operators, computation graphs, memory optimisation, model interoperability and community maintenance.
86.6% lower Mask R-CNN CPU inference memory with graph-level memory reuse, compared with the same framework without graph optimisation.
Explore OwlA research project on timely inference for driving models. I designed a common model-adaptation and replay-evaluation framework and used controlled experiments to investigate performance.
17.9% lower median per-frame replay service time in an offline development benchmark, with output-consistency checks.
Explore OverDriveA legal knowledge system combining hybrid retrieval, entity knowledge and MCP tools. Designed, developed and deployed on AWS, with feedback from a pilot at a leading Chinese law firm.
Incremental knowledge maintenance, source-checked quotations and asynchronous processing; server costs reduced by approximately 67%.
Explore KnowledgeBaseAt KIT, I developed a real-time water-level prediction module for a Rhine inland-port monitoring system with Covadonga; the module is in live operation.
Earlier work includes the Databox personal-data platform, distributed and federated learning, and multi-agent reinforcement learning for crowdsensing. See research for the publications.
Course implementations and independent study are collected separately in my learning notes.