Software engineer building at the systems level — GPU kernels, C++ performance work, and AI-assisted developer tooling.

Currently contributing GPU kernels to open-source ML infrastructure (llama.cpp) while building RAAG, an architectural analytics platform that scopes AI refactoring to exactly what a codebase change can reach.

llama.cpp — GPU inference engine (30k+ ⭐)

• Implemented a CUDA kernel for 1D pooling (avg/max), closing a backend coverage gap. Verified across 216 test cases on 2x NVIDIA T4 GPUs. Merged into master by the project's lead maintainer. → PR #27573
• Metal backend contribution — causal attention masking op. → PR #27197 (in review)

AI-Powered Architectural Analytics Platform — parses a codebase, quantifies its coupling/cohesion, and scopes AI refactoring suggestions to exactly what a change can reach (blast-radius, not text similarity).

• 3.69x parsing speedup via a std::jthread pool over 579 real C++ source files (349 → 1290 files/sec)
• 86% test coverage across 307 tests
• Three-engine, three-language pipeline: C++20 extraction (Tree-sitter) → Python dependency-graph analytics (NetworkX) → GraphRAG-scoped retrieval and reasoning (Qdrant + Claude)
• CI-gated: blocks a pull request automatically if it pushes a core module's instability past threshold
• Ships as a CLI and a VS Code extension

📺 Demo walkthrough · 📦 Repo

Writing about it:

• Building a Parallel C++ Source Parser: jthread, stop_token, and the Deadlock I Didn't See Coming
• I Ran a Coupling Analyzer on nlohmann/json and fmt. It Found a Class Doing 55 Jobs
• I Built an AI Refactoring Tool That Can't See More Code Than the Dependency Graph Allows

C++20 CUDA Python CMake Tree-sitter GraphRAG Docker

LinkedIn · Hashnode