Workload-specialized allocator optimization
Currently in progress

An ongoing project that extends a custom memory allocator into a real Tree-sitter parsing workload. The goal is to analyze allocation behavior, compare performance against glibc malloc, and use benchmark data to guide workload-specific allocator optimization.

Custom dynamic memory allocator in C
malloc/free/calloc/realloc implemented from scratch.

Implements 16-byte alignment, 4-byte headers, boundary tags, block splitting, coalescing, free-list management, and heap invariant checking. This project helped me understand how dynamic memory allocation works below the standard library.

Zoomable codebase exploration tool
“Google Earth for your codebase.”

Parses repositories with Tree-sitter, visualizes code structure across file/class/function/code layers, estimates complexity, and recommends code reading order using PageRank and Betweenness Centrality.

Secure split learning system
Privacy-preserving collaborative training.

Dynamically builds PyTorch models from JSON configs, validates model configs inside a WASM sandbox, and isolates node-side processes using Linux namespace and seccomp-bpf to reduce security risks.

Real-time Wikimedia edit stream analyzer
Detects high-tension Wikipedia articles.

Ingests live enwiki edit events, maintains Redis hot-tier and SQLite cold-tier storage, and detects signals such as 3RR, edit velocity spikes, and editor conflicts.

Runtime · Memory · Systems Performance

I am currently focusing on allocator internals, Linux memory behavior, Tree-sitter workload analysis, and runtime-level performance bottlenecks.