A hierarchical plan-first agent (Goal → Parts → Steps) that decomposes high-level tasks into verifiable sub-goals, cutting redundant LLM calls by 60% vs. pure ReAct by executing deterministic steps open-loop. Self-healing loop catches mistakes and retries with better insight. Compiled to a downloadable desktop app via Tauri.
Stack: Python, TypeScript, C++, WebKit, Gemini API, Tauri
Tags: Agents, LLM, Desktop
A high-performance C++/CUDA linear layer kernel fusing INT4/INT8 weight de-quantization with FP16 tensor core matmul to accelerate LLM inference. Cut L1 cache misses 50% by packing low-precision weights into integer types with fast bit-shift unpacking, and eliminated shared memory bank conflicts via memory padding.
Stack: C++, CUDA, Git
Tags: CUDA, Performance, ML Infra
A full-stack hardware emulator replicating the CHIP-8 interpreter: 4KB memory layout, 16 general-purpose registers, hardware stack, and a fetch-decode-execute cycle handling all 35 opcodes. 60Hz pixels crate based rendering pipeline. Supports Superchip, CHIP-8 and Atari game ROMs, and tested with Space Invaders, and various other public IP games. Can set clockspeeds and FPS dynamically during build.
First time making something in Rust, and actually learn how to use macros, defining enums, shortcuts while coding, etc. CHIP-8 build was hardcoded, Superchip implementation was then added later with the help of AI.
Stack: Rust
Tags: Emulation, Systems
Finetuned a Qwen 3.5 9B model for autonomous web navigation with specialized tool-calling loops over real-time financial market environments. Deployed via a high-performance serving engine with optimized KV-cache management, and quantized to 4-bit (AWQ/GPTQ) for over 60% memory reduction at high-density edge deployment.
Stack: Python, PyTorch, TensorRT-LLM, Docker
Tags: ML, LLM, Quantization
This site — a portfolio that behaves like a terminal by default, with a fixed set of page commands falling through to a real backend Linux sandbox (per-session Firecracker VMs, real gcc/rustc/python3, encrypted snapshot save/restore) for anything else, plus a modern "Easy Mode" hiding underneath.
Stack: React, Vite, Rust, Node.js, Fly.io
Tags: React, Infra