Most of what I ship is just stuff I needed and refused to pay someone else to half solve for me. Some of it's for fun. Some is a proof of concept that got out of hand. Some is me learning because that's less embarrassing than it sounds. A few exist for no reason I can fully defend. All of it meant something at the time. All of it is my stubborn belief that DIY is still the only way out.
A bootstrapped, solo-dev AI workstation for lyric writing and music production now a multi-app platform spanning 3 deployments, 3 separate codebases, and 9 proprietary model variants. Grew from a personal tool to daily active users and paying subscribers with zero advertising.
Platform Surfaces:
| App | Stack | Status |
|---|---|---|
| vrsa.app | React + Vite + Capacitor 8 | Live |
| studio.vrsa.app | Next.js 16 + TypeScript + Docker | Live |
| generate.vrsa.app | Next.js + React + Zustand + WaveSurfer.js | WIP |
View Platform Architecture Diagram
graph TB
subgraph USER [" "]
direction LR
Web["Web"]
PWA["PWA"]
Android["Android"]
end
Web & PWA & Android --> A
subgraph APPS ["Three Deployments"]
A["<b>vrsa.app</b> React · Vite · Capacitor · 32 routes · 50 components · 9 hooks · Ghostwriter · Canvas · Audio Analyzer · Suno Bridge · Album Art · Rights Mgmt"]
B["<b>studio.vrsa.app</b> Next.js · TypeScript · Tone.js · Browser DAW — 16 tracks · 64 clips · Piano Roll · Drum Sequencer · Waveform Editor · 1,300 lines of synth presets · MIDI I/O"]
C["<b>generate.vrsa.app</b> Next.js · TypeScript · Zustand · 11 pages · 8 stores · ~120 components · 7 gen modes · AI Radio · AI DJ · LoRA Training · WaveSurfer.js"]
end
Web --> B & C
subgraph AUTH ["Shared Auth Layer"]
D[("Supabase · .vrsa.app cookies · RLS · Edge Functions · RPCs · Storage")]
end
A & B & C --> D
subgraph GPU ["Modal Backend — H100 GPU"]
E["Async Spawn/Poll · 600s Timeout · Persistent Weight Volume"]
E --> F["<b>6 DiT Models</b> v1.5 · v1.5-stream · v1.5-craft · v1.5-core · v1.5-shift3 · v1.5-shift1"]
E --> G["<b>3 LM Models</b> pulse-nano 0.6B · pulse 1.7B · pulse-pro 4B"]
end
D --> E
subgraph LLM ["Multi-Provider LLM Routing"]
H["GPT · Claude · DeepSeek · Gemini · Grok · Kimi K2 · MiniMax · Mistral"]
end
A --> H
subgraph MEDIA ["Media Generation"]
I["Flux 1.1 Pro · Flux 2 · Sora via Azure OpenAI"]
end
A --> I
subgraph INFRA ["Infrastructure"]
J["107 API Endpoints · Vercel · Docker · Sentry · Honeybadger · Amplitude · Intercom · Cloudflare Turnstile · FingerprintJS · Crowdin i18n"]
end
style USER fill:transparent,stroke:#58A6FF,stroke-width:2px,color:#58A6FF
style APPS fill:transparent,stroke:#8B5CF6,stroke-width:2px,color:#8B5CF6
style AUTH fill:transparent,stroke:#3ECF8E,stroke-width:2px,color:#3ECF8E
style GPU fill:transparent,stroke:#F97316,stroke-width:2px,color:#F97316
style LLM fill:transparent,stroke:#EC4899,stroke-width:2px,color:#EC4899
style MEDIA fill:transparent,stroke:#EAB308,stroke-width:2px,color:#EAB308
style INFRA fill:transparent,stroke:#6B7280,stroke-width:2px,color:#6B7280
View VRSA v1.5 AI Music Generation Engine Specs
Custom cloud inference backend running 6 DiT plus 3 LM model variants on H100 GPUs, all served from a persistent weight volume with async spawn/poll architecture:
| VRSA Name | Type | Description |
|---|---|---|
| vrsa v1.5 | DiT (default) | Turbo — balanced speed + quality |
| vrsa v1.5-stream | DiT | Continuous generation mode |
| vrsa v1.5-craft | DiT | SFT — higher fidelity |
| vrsa v1.5-core | DiT | Base model |
| vrsa v1.5-shift3 / shift1 | DiT | Flow-shift variants |
| vrsa-pulse-nano | LM 0.6B | Fast token-level generation |
| vrsa-pulse | LM 1.7B | Balanced LM |
| vrsa-pulse-pro | LM 4B (default) | Full reasoning + CoT metadata |
|
|
|
View 30+ Page Modules Architecture
| Module | Description |
|---|---|
| Ghostwriter | Chat-based lyric generation with session memory, multi-model A/B mode, take history, and album-context awareness |
| Canvas | Notion-inspired editor with inline AI edits, syllable counter, rhyme heatmap, and MP3 transcription |
| Suno Bridge | Chrome + Firefox extensions that capture auth tokens and route lyrics directly to the Suno with version picker (v4 → v5) |
| Audio Analyzer | BPM + key/scale detection via custom VM audio engine running Essentia.js WASM |
| Album Art | Flux 1.1 Pro, Flux 2, and Sora via Azure OpenAI deployments |
| Multi-model routing | GPT, Claude, DeepSeek, Gemini, Grok, Kimi K2, MiniMax, Mistral across OpenAI, Bedrock, OpenRouter, and Google APIs |
| Rights Management | Proof-of-creation PDF certificate generator |
| Workspace | Full library, project, and album management |
| Admin Panel | Platform administration and stats CMS |
| Studio Pass | Subscription and billing management |
| Mobile | Capacitor 8 Android build + full PWA support |
| Infrastructure | |
|---|---|
| Auth | Cross-subdomain cookies scoped to .vrsa.app, shared across all three apps |
| Compute | Modal async generation pipeline, H100 GPU, 600s timeout, persistent weights volume |
| API | 107 endpoints mapped and routed |
| DevOps | Crowdin i18n workflow, Docker + docker-compose, multi-project deployment |
- Pentesting Explanations
A high-quality supervised fine-tuning dataset for penetration testing expertise and red team tradecraft built to improve model reasoning on offensive security. Matures toward novel vulnerability research and zero-day reasoning. - Linux ASM Pairs
A dataset of disassembled Linux kernel functions paired with structured natural-language explanations, built from the Linux Commits Dataset (Zenodo 10654193). - Genius Lyrics Dataset
3.17M cleaned and deduplicated English song lyrics sourced from Genius (title, artist, genre tag, year). MIT licensed. Built for training, fine-tuning, and music analysis at scale.
A curated list of 100+ AI-powered audio tools across 16 categories including music creation, voice cloning, stem separation, TTS, transcription, sound detection, and more.
- Website to Markdown Exporter
A userscript that exports the fully rendered text of any webpage to a clean, well-structured Markdown file. Works on the live DOM, meaning JavaScript-rendered content is captured correctly. - eDEX-UI (Security Patched)
The originaleDEX-UI(a well-known sci-fi terminal emulator) was archived with an unresolved security vulnerability. I forked it, patched the flaw, modernized the codebase, removed deprecated dependencies, and cut a clean release. - CorgiDB Contribution
Contributed a helper function to automate efficient SQL datatype selection for pandas DataFrames reducing manual overhead and ensuring optimal storage across schema generation workflows.
| Tool | Description | Open Notebook |
|---|---|---|
| HeartLib Script | Optimized scripts for running HeartMuLa music generation with maximum performance on NVIDIA GPUs (especially A100). |  |
| Ollama Stack | Full Ollama stack running on Google Colab with Gradio as a UI. Spin up a local LLM in your browser with zero local setup. | |
| Wan 2.2 Video | Plug-and-play Colab notebook for Wan 2.2 — an advanced image-to-video model — stripped down to just work. | |
Certifications
t̵̨̧̝̯̺͚̖̬̲̜͓̩̟̼͈̋͛̑̒̔̀͊́͜͝ͅh̴̨̦̹̖̟̺̬̫̩̲̦͉̪̩̹͓̲̩͐̇̈́́̆͋̓̽͋̚e̶̢̛̮̟̝̗͙͙̟̣̤͓̙̻̜͖̅́̔̍͛̆̊̏̉̑̓̋͂̽̏͗̓̄͛̽̋̈́͘͘͘͠͝͠ę̵̨̧̨̨̱̙̱͈̞͙̮̭̫̼̝͇͓͉̦̯͍̪̣̰̠̹͚͂̽̄̆̒̋̍̀̇̊̌̑͗̍̀̽̈́̄̿̓̋̔̿̔̈́͘l̷̡̝̲̻̖̬̹͈̈́́̐̀̈̍̌̂͋̑͆̌͐̒̀̔̒̓̄̃̄̿̓́̈̐͝͝͝͠ḑ̸̢̬͕̺̹̥̠̯̳͍͖̰̀̀̀͒̋̋͜ë̸̡͇̺́́̍̾̃͗͐̐̀̑́̿͆̉̕͝͝͠͝͠r̵̡̧̲̰̥̘̮̭͕̉̓̽̈̍̄̍̾͂̕e̵̡̡̡̟̱͇̯̼̰̞͚̻̺̠͕̠̤̽̓̒̀͗͠ṃ̶̨̞̈́̈̏͗̋̓͂̏̎̍̋͌̾̚̚ơ̵̩̱̮͖̭̤̯̝̹̼̎͑̍́̄͐̈́ͅ