Building Software for Independence & Performance
I design and build desktop applications, neural network architectures, and developer tools engineered for local hardware execution. As neural models become more efficient and consumer GPUs gain substantial tensor acceleration, software should empower users with full data ownership, predictable latency, and independence from cloud subscription tiers.
My work spans local speech recognition studio workflows (SubtitleGo), deep reinforcement learning engines inspired by AlphaZero (OceanGo), and self-hosted collaborative systems (Bathyal).
Algorithmic Exploration & Local Intelligence
My deep interest in artificial intelligence extends beyond consuming high-level cloud APIs into the core mathematical mechanics of neural architectures and decision trees:
- Reinforcement Learning & Tree Search: Recreating AlphaGo and AlphaZero architectures on consumer hardware through deep Residual Networks (ResNet) with dual policy/value heads combined with asynchronous Monte Carlo Tree Search (MCTS) and Dirichlet noise exploration.
- In-Process Neural Inference: Integrating state-of-the-art speech and vision models (such as Qwen3-ASR and Whisper) directly inside native desktop process memory via PyTorch, ONNX Runtime, and Apple Metal Performance Shaders (MPS).
- Local LLM Experimentation: Profiling, fine-tuning, and running open-weight language models locally to test parameter efficiency, context window utilization, and agentic workflows without telemetry.
Core Interests & Creative Domains
AI & Local LLMs
Experimenting with offline model quantization, fine-tuning datasets, and exploring generative AI mechanics directly on edge hardware.
Game Development
Exploring interactive simulation, game mechanics, 2D/3D physics engines, and procedural systems.
Photography
Observing and capturing moments, natural lighting, geometry, and compositional balance in the physical world.
Engineering Philosophy
Every open-source project in this ecosystem adheres to four foundational architectural principles:
Local-First & Private
All compute, audio decoding, and model weights run strictly inside local process memory. No telemetry and no remote network dependencies.
Direct In-Process Runtimes
Avoid heavy web servers or microservice containers for desktop workflows. Leverage PyTorch, ONNX, and TensorRT directly.
Cross-Platform Acceleration
Deep optimization across NVIDIA CUDA, Apple Metal (MPS), and multi-threaded CPU fallbacks for broad accessibility.
Standardized Distribution
Provide pre-built standalone ZIP archives for one-click usage alongside clear developer source build instructions.
Core Toolchain & Technologies
Technologies and frameworks actively utilized across my projects:
Open Source Projects
Explore dedicated architecture breakdowns and source repositories across my projects:
SubtitleGo
Local-first speech recognition and subtitle editing studio powered by PySide6 and Qwen3-ASR (1.7B).
OceanGo
AlphaZero-style Go AI using 19-block ResNet, dual policy/value heads, and batched MCTS with virtual loss.
Bathyal
Lightweight, self-hosted project management system with Kanban boards, time tracking, notifications, and automations.