AI-powered legal document processing platform. Solo-built the entire backend: FastAPI, async SQLAlchemy, PostgreSQL on GCP serverless. 7-phase document pipeline (500+ PDFs in <10 min), multi-agent contract drafting with 19 templates, LLM infrastructure with prompt caching and provider abstraction.
Ported the SCAMP matrix profiling algorithm to the Cerebras CS-2 Wafer-Scale Engine. Benchmarked non-AI workloads on massively parallel hardware, measuring PCIe bandwidth and FLOPS against GPU/CPU baselines. Tech: Cerebras CSL, Python, SLURM.
Optimised Jacobi iterations with NUMA-aware AVX-512 intrinsics across 1000+ AMD EPYC nodes. Benchmarked with MPI/OpenMP hybrid parallelism on LRZ's HPC infrastructure. Tech: C++, AVX-512, MPI, OpenMP, SLURM.
Custom reliable transport protocol built on UDP with MTU negotiation, packet-drop recovery, and checksum validation. Designed for lossy networks. Tech: Rust, Tokio, async I/O.
Contributed to the default music player for the GNOME desktop environment. 14+ commits to the upstream project covering UI features and backend improvements.