Reason without language.
I build AI frameworks that reason without LLMs, learn without backpropagation, and run on commodity hardware.
AI Engineer and systems programmer building deterministic systems on commodity hardware.
I build AI frameworks that reason without LLMs and learn without backpropagation — designed to run on commodity CPUs and edge devices instead of GPU clusters.
My work spans Rust (vAGI-2, NSE, pse-engine), Python (AOT-Compiler, RE-super-agent, bedrock-obfuscator), Go agents, and C++ graphics hooking — with LLVM IR compilation and WebGPU along the way.
Core thesis: correctness over confidence, efficiency over scale, verification over guessing.
- 01 CPU-first AI (no GPU clusters)
- 02 Compilers & LLVM IR
- 03 Reverse engineering & security
// Rust · Python · Go · C++ · TypeScript
Work
Selected systems, compilers, and AI frameworks — all open source.
CPU-first cognitive architecture — BitNet ternary weights + Hamiltonian dynamics + symbolic reasoning.
Experimental Rust workspace for CPU-friendly language modeling. Ternary weights {-1,0,+1}, SIMD kernels, and supporting crates for simulation, symbolic math, memory, and reasoning.
Python source obfuscator with a native C-interpreted bytecode VM.
Monolithic obfuscator for CPython 3.10–3.12. Emits obfuscated .py extremely hard for humans and decompilers to read, with an optional native C bytecode VM and FastAPI playground.
Super agent for reverse engineering — 5-domain MCP + multi-specialist orchestration.
Hybrid RE agent: 5 domain MCP servers, Python multi-specialist core, dynamic workflow engine, safety/isolation layer. 44 tools across static/dynamic/symbolic/deobfuscation/malware domains wrapping Ghidra, radare2, angr, Frida, Qiling.
AOT compiler pipeline compiling a typed subset of Python to optimized LLVM IR.
AOTC compiles Python directly to optimized LLVM IR with a Zero-Copy Data Bridge letting native functions operate directly on external buffers without copying.
Neuro-Sparse Engine — run sparse LLMs on CPU/Edge without GPU clusters.
Rust research framework running models in sparse + quantized form. ZSTM transmutation, RIE/HNSW routing, LLER AVX2 kernels. End-to-end pipeline from training to sparse inference and evaluation.
Physical & shader engine for the web — Rust → WASM, WebGPU + WebGL2.
Reusable Rust + WebAssembly physics and shader engine supporting modern (WebGPU) and legacy (WebGL2) browsers via wgpu. GPU compute broad-phase + CPU narrow-phase + Rust solver, single shared GPU storage buffer for instanced rendering.
Principles
The rules I build by.
Correctness > Confidence
If the code is wrong, it fails. Confidence is a proxy, not a metric.
Efficiency > Scale
Run on commodity hardware. Scale is a side effect, not a goal.
Verification > Guessing
Never trust a promise. Only trust a checksum.
Evidence
Public artifacts and what they prove.