Guiding Quote
"I am not designed to come second or third. I am designed to win."
Technical Arsenal
Core systems engineering, ML research, and acceleration tooling.
Core & Systems
C++
Qt
WinAPI
Linux
C#
Machine Learning & AI
Transformers
ML (General)
GANs
TensorFlow
Diffusion
Hardware & Acceleration
HIP/ROCm
TPUs
CUDA
Timeline Portfolio
Structured in multi-year arcs, ordered newest to oldest. Each entry is either a project case study, a research case study, or a milestone that marks a career-shaping shift.
Research Card
Project Card
Milestone
Neo-Hybrid Arc
2026-present
Research, product, and business reinforce each other in a single operating loop.
Research
Brontes
Waveform-domain enhancement model built to repair neural codec artifacts while upsampling degraded 24 kHz speech to 48 kHz output. It follows a synthesis-first design to rebuild detail instead of preserving artifact patterns.
- Synthesis-first stack with selective deep skips, temporal compression, and LSTM bottleneck
- Gated decoder fusion with adversarial objectives in time and spectral domains
- Stabilized training through discriminator representation changes
Research
VITS EVOlution (VITS EVO)
Optimized evolution of the VITS text-to-speech architecture focused on extreme real-time performance with practical audio quality. Current builds reach about 300x real-time on an NVIDIA A6000 and run in TensorVox through ONNX + DirectML on Windows 10+ DX12 GPUs.
- Architecture and inference optimizations targeted at low-latency, real-time use
- Model is private, but production integration is documented in TensorVox
Milestone
Research-to-Product Validation in TensorVox
Integrated VITS EVO into TensorVox with ONNX + DirectML, proving native Windows 10+ GPU inference across DX12 hardware and demonstrating a clear path from model research to user-facing delivery.
Pure Research Arc
2023-2025
Deep neural audio exploration with frequent experimental iteration and publication.
Research
Echolancer
Text-to-speech model trained on concatenated text-audio tokens using a decoder-only LLM-style architecture. Uses a 50 tok/s codebook for efficient audio language modeling with high-fidelity output.
- 50 tok/s codebook design for efficient audio token pacing
- Single-stream sequence framing for language-audio alignment experiments
Research
MQGAN / MusicLSTM
Experimental audio tokenization built around a spectrogram codec, with a ResNet compressor and U-Net refiner for cleaner reconstruction. Designed to keep tokenized workflows compact while improving output quality.
- Spectrogram-first codec framing for music-oriented generation tests
- Compressor-refiner split for better detail recovery in outputs
Milestone
ROCm Star Recognition
Earned ROCm Star recognition through hands-on work with AMD acceleration hardware and tooling in real model training contexts.
Research
ReLUGT Activation
Project
WinDiffusion
Lightweight C++/Qt frontend for Stable Diffusion with no Python dependency for end users. Supports txt2img, img2img, and inpainting in a native Windows workflow.
- Supports txt2img, img2img, and inpainting in one native interface
- Built for simpler onboarding and repeatable local execution
Milestone
Startup Transition and Deeper Model Work
Joined a startup context and shifted into deeper architecture and training-detail work, tightening the loop between research decisions and shipped outcomes.
Hybrid Arc
2020-2023
Software engineering and ML research overlapped to produce practical model tooling.
Research
Tacotron 2 Conv Attention
Convolutional Attention Consistency approach for stabilizing Tacotron 2 attention. Built to reduce alignment failures and improve training reliability.
- Targets known instability patterns in sequence-to-sequence speech synthesis
- Framed as practical experimentation with reproducible writeups
Project
TensorVox
High-performance local TTS application written in C++/Qt with support for both PyTorch and TensorFlow models. Widely used for freelancing deliverables and model deployment workflows.
- Served as the execution backbone for client and freelancing deliverables
- Bridged model experimentation with practical desktop UX
Milestone
Became TensorFlowTTS Maintainer
Contributed the first 44.1KHz open-source pretrained model in the repository, expanding practical high-fidelity speech options for the community.
Milestone
Added C++ Support to TensorFlowTTS
Implemented C++ export and runtime usage through the TensorFlow SavedModel C API, broadening deployment options beyond Python-only workflows.
WinDev Arc
2015-2020
Early system-builder years focused on complete desktop tools and shipping discipline.
Project
ZMapCharter
Statistical map charting tool built with C++ and SFML. Supports custom shapefiles and CSV/XLSX imports for end-to-end data-to-visualization workflows.
- Supports custom shapefiles and CSV/XLSX data workflows
- Built for practical analysis tasks instead of toy demos
Project
ZDEditorRS
Modern C++ rewrite of the SuperPower 2 database modding tool. Became the community standard for modding with a maintainable long-term codebase.
- Focused on reliability for power users managing large game data edits
- Demonstrated maintainable architecture in an active modding ecosystem
Milestone
Video Game Modding Phase
Intensive C++, Qt, and WinAPI work through community tooling built strong systems instincts, fast iteration habits, and direct feedback loops with real users.
Milestone
Desktop Engineering Foundations
Early software years moved from Visual Basic and C# roots into C++-first development, setting up the systems depth that later supported ML and research work.