Skip to content
View theelderemo's full-sized avatar
:shipit:
Living in a constant merge conflict
:shipit:
Living in a constant merge conflict

Block or report theelderemo

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
theelderemo/readme.md
The Elder Emo
DOI Cybersecurity Student US Army Vet TS Clearance

Army veteran, cybersecurity student, and the type to fork an archived project rather than file an issue and wait.


VRSA

vrsa.app - an AI workstation for songwriters, covering lyric writing, production, and generation. Bootstrapped, solo-built, zero advertising, and it has daily active users and paying subscribers.

Three deployments and three codebases that behave like one product:

Surface Stack Status
vrsa.app React · Vite · Capacitor 8 Live
studio.vrsa.app Next.js 16 · TypeScript · Tone.js · Docker Live
generate.vrsa.app Next.js · Zustand · WaveSurfer.js WIP

The decisions I'd actually defend in review:

  • One session, three apps. Auth is a single Supabase cookie scoped to .vrsa.app, so signing into any surface signs you into all of them. RLS does the authorization; the apps don't hand tokens to each other.
  • Generation is spawn-and-poll, not request-response. Music generation on an H100 outlives any reasonable HTTP timeout, so jobs are spawned on Modal and polled to a 600s ceiling. Weights sit on a persistent volume instead of cold loading nine model variants per request.
  • The LLM layer is a router, not a vendor. GPT, Claude, Gemini, DeepSeek, Grok, Kimi K2, MiniMax, and Mistral across Azure, OpenAI, Bedrock, OpenRouter, and Google.
  • Audio analysis stays in the pipeline, not a third-party API. BPM and key/scale detection run through a custom audio engine on Essentia.js compiled to WASM.
  • Suno Bridge ships as browser extensions. Chrome and Firefox builds capture auth and route lyrics straight into Suno with a version picker.
Platform data flow
flowchart TD
    U["Web · PWA · Android"] --> AUTH

    AUTH[("Supabase - one cookie scoped to .vrsa.app<br/>session · RLS · storage · edge functions")]

    AUTH --> APP{Surface}

    APP -->|"lyrics · canvas · art"| A["<b>vrsa.app</b>"]
    APP -->|"multitrack DAW"| B["<b>studio.vrsa.app</b>"]
    APP -->|"music generation"| C["<b>generate.vrsa.app</b>"]

    A --> LLM["LLM router<br/>GPT · Claude · Gemini · DeepSeek<br/>Grok · Kimi K2 · MiniMax · Mistral"]
    A --> IMG["Flux 1.1 Pro · Flux 2 · Sora<br/>via Azure OpenAI"]
    A --> EXT["Suno Bridge<br/>Chrome + Firefox extensions"]

    B --> TONE["Tone.js engine<br/>piano roll · sequencer · MIDI I/O"]

    C --> SPAWN["Modal - spawn job"]
    SPAWN --> POLL["poll until ready<br/>600s ceiling"]
    POLL --> GPU["H100 · persistent weight volume<br/>6 DiT + 3 LM variants"]
    GPU --> OUT["mp3 · flac · wav · opus · aac"]

    LLM & IMG & TONE & OUT --> STORE[("Supabase Storage")]
    STORE --> U

    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 IMG fill:transparent,stroke:#EAB308,stroke-width:2px,color:#EAB308
    style STORE fill:transparent,stroke:#3ECF8E,stroke-width:2px,color:#3ECF8E
Loading
v1.5 generation engine - model variants & controls

Custom cloud inference backend serving six DiT and three LM variants from a shared weight volume:

Variant Type Role
vrsa v1.5 DiT (default) Turbo - balanced speed and quality
vrsa v1.5-stream DiT Continuous generation
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
vrsa-pulse-pro LM 4B (default) Full reasoning + CoT metadata

Exposed controls: BPM, key/scale, time signature, guidance, ODE/SDE sampling, CFG Task types: text-to-music, music-to-music, cover, stem extract, repainting, build, complete Output: mp3, flac, wav, wav32, opus, aac


Security & Research

StreamRecorder IDOR Disclosure Two API endpoints on streamrecorder.io returned private recording direct stream/download URLs to any authenticated user. Both accepted user supplied IDs without checking whether the caller owned the object. Written up as CWE-639 with a risk breakdown, reproducible proof of concept, and a disclosure trail.

Pentesting Explanations A 5,906-row supervised fine-tuning dataset for adversarial reasoning. Most security datasets teach recognition - give the model a technique name, get back its description, which produces models that don't context through or reason the moment they meet an unfamiliar codebase. Every row here carries a think trace written from the operator's seat, option by option. Apache 2.0, shipped in SFT / GRPO / PRM formats, sharded by HackTricks and MITRE ATT&CK.

eDEX-UI (security patched) The original sci-fi terminal emulator was archived with an unresolved vulnerability. I forked it, patched the flaw, and did the dependency archaeology: Electron 12 to 37 across 25 major versions, the scoped @xterm packages, node-pty 0.10 to 1.0, ws 7 to 8, plus a clean release and a security policy.


Systems & Tools

tunesort - Rust · GTK4 · GStreamer A keyboard-driven triage tool for people with more music than patience. Native GTK4, no Electron, one small binary. Waveform peaks are pulled out through GStreamer and drawn with Cairo so you can scrub by dragging the wave; MIDI files play through GStreamer's synth pipeline and get a note-density graph in place of a waveform. Deletes go to the freedesktop trash and get pulled back out on undo - multi-level, for moves as well as deletes - because I do this half-asleep and don't trust myself with unlink. CI, release automation, and .deb packaging with a desktop entry.

hawktui - Python · Textual · yt-dlp A TUI that sits in the background watching your clipboard and hands any URL you copy straight to yt-dlp. Quality and format selection, audio-only mode, live progress, persisted config. Tested and shipped through a release pipeline. Its README is written in a rather different register than this one.

Website to Markdown Exporter - Userscript Exports the fully rendered text of any page to clean Markdown. It walks the live DOM rather than the source, so JavaScript-rendered content actually makes it out, and it resolves every link and image to an absolute URL so the file still works away from its origin. On GreasyFork for Violentmonkey and Tampermonkey.

Smoked - Firefox userChrome.css A dark, glassy Firefox theme with a one line toolbar and a zen mode that collapses the entire browser UI to thin strip, revealing on hover or focus, with tab positions still readable as colored slivers. Frosted glass panels via backdrop-filter, themed about: pages.

CorgiDB - Contribution Added a helper that automates SQL datatype selection for pandas DataFrames, cutting the manual pass out of schema generation and keeping column storage sane.


ML & Data

Linux ASM Pairs Disassembled Linux kernel functions paired with structured natural-language explanations, built from the Linux Commits Dataset (Zenodo 10654193).

Genius Lyrics Dataset 3.17M cleaned, deduplicated English song lyrics with title, artist, genre tag, and year. MIT licensed, built for training and analysis at scale.

AI Audio Tools A curated index of 100+ AI audio tools across 16 categories - music creation, voice cloning, stem separation, TTS, transcription, sound detection.

Colab Stacks

Notebook What it does Open
HeartLib HeartMuLa music generation tuned for NVIDIA GPUs, A100 in particular Open In Colab
Ollama Stack Full Ollama stack with a Gradio UI - a local LLM in your browser, zero local setup Open In Colab
Wan 2.2 Video Image-to-video generation, stripped down to just work Open In Colab

Certifications

Google Cybersecurity Professional Certificate V2 Google Project Management Professional Certificate Introduction to Cybersecurity LFC108: Cybersecurity Essentials Networking Basics Network Addressing and Basic Troubleshooting Introduction to Cybersecurity Tools and Cyberattacks V3 Cybersecurity Roles, Processes & Operating System Security Project Management Essentials Cyber Threat Management Meta Data Analyst Professional Certificate Cybersecurity Essentials IBM Cybersecurity Analyst Assessment


t̵̨̧̝̯̺͚̖̬̲̜͓̩̟̼͈̋͛̑̒̔̀͊́͜͝ͅh̴̨̦̹̖̟̺̬̫̩̲̦͉̪̩̹͓̲̩͐̇̈́́̆͋̓̽͋̚e̶̢̛̮̟̝̗͙͙̟̣̤͓̙̻̜͖̅́̔̍͛̆̊̏̉̑̓̋͂̽̏͗̓̄͛̽̋̈́͘͘͘͠͝͠ę̵̨̧̨̨̱̙̱͈̞͙̮̭̫̼̝͇͓͉̦̯͍̪̣̰̠̹͚͂̽̄̆̒̋̍̀̇̊̌̑͗̍̀̽̈́̄̿̓̋̔̿̔̈́͘l̷̡̝̲̻̖̬̹͈̈́́̐̀̈̍̌̂͋̑͆̌͐̒̀̔̒̓̄̃̄̿̓́̈̐͝͝͝͠ḑ̸̢̬͕̺̹̥̠̯̳͍͖̰̀̀̀͒̋̋͜ë̸̡͇̺́́̍̾̃͗͐̐̀̑́̿͆̉̕͝͝͠͝͠r̵̡̧̲̰̥̘̮̭͕̉̓̽̈̍̄̍̾͂̕e̵̡̡̡̟̱͇̯̼̰̞͚̻̺̠͕̠̤̽̓̒̀͗͠ṃ̶̨̞̈́̈̏͗̋̓͂̏̎̍̋͌̾̚̚ơ̵̩̱̮͖̭̤̯̝̹̼̎͑̍́̄͐̈́ͅ

Pinned Loading

  1. tunesort tunesort Public

    A dark, keyboard driven tool for triaging large music libraries: shuffle, audition, trash, and quick move tracks. Built with Rust

    Rust 1

  2. hawktui hawktui Public

    Hawk tuah on that clipboard and spit straight to disk.

    Python 1

  3. StreamRecorder-IDOR-Disclosure StreamRecorder-IDOR-Disclosure Public

    Technical disclosure of a critical IDOR vulnerability (CWE-639) in streamrecorder.io allowing unauthorized access to premium video content.

  4. eDEX-UI-security-patched eDEX-UI-security-patched Public

    Forked from GitSquared/edex-ui

    A cross-platform, customizable science fiction terminal emulator with advanced monitoring & touchscreen support.

    JavaScript 57 4

  5. wan2.2-google-colab wan2.2-google-colab Public

    This allows you to utilize Google Colab (or other notebooks) to run wan2.2 image to video

    Jupyter Notebook 48 16