This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

TL;DR

• Built for Nathan: best friend of 20+ years, musician behind Flies Like Robots, not a marketing guy.
• Self-hosted Gemma 4 watches and listens to his video, then drafts the YouTube metadata and Bandcamp copy in his voice, and picks a hook for a vertical Short.
• He edits or approves. Soundboard remembers the edit, uploads as private, and reads it back from YouTube before calling it verified.
• Code • 60 s demo • Deployed site

What I Built

Nathan makes the music and the videos. He's been one of my best friends for more than 20 years, so I can say this with some authority: he is a musician, and asking him to also be in marketing is seriously pushing it.

Before it drafts anything, Gemma reads Nathan's 30 latest videos and 10 thumbnails, then proposes a brand guide with keep, fix, and drop lists for him to edit or approve. That approved guide takes priority over patterns in his older uploads, because reading his channel shouldn't mean repeating everything he's trying to improve. I want the copy to sound like Nathan, and “write me a YouTube title” gives the model nothing to work with.

Soundboard covers the gap between a finished video and somebody finding it. He uploads the video and types the song title. Gemma drafts the YouTube title, description, and tags, grounded in what's reaching listeners in that genre and written the way Nathan writes. It drafts the Bandcamp About text and credits too, with a Copy button for each, because Bandcamp has no upload API and won't be framed. And it picks the hook for a vertical Short, which ffmpeg cuts to 9:16. He edits, re-runs, or approves. Every edit becomes evidence for the next draft.

I can't promise an algorithm will get his music heard. I can take the busywork off him.

His music: YouTube • Bandcamp.

Demo

One sample, start to finish: upload, Gemma's draft, an edit, approval, and the private upload verified against YouTube. To run it locally, follow the README's GCP setup, pull the Gemma 4 GGUF for llama-server, then make dev.

Sign in with username anchildress1demo@gmail.com and password xAf(nH*2Qf1T to access the YouTube sandbox. You can also sign into YouTube directly to validate your upload.

Deployed at https://soundboard.anchildress1.dev.

Code

⚖️ MIT. Gemma's weights ship separately under Apache 2.0.

The requirement: if Nathan fixes a title once, he shouldn't fix it the same way twice. For this build, I pass his saved edits into the next prompt instead of maintaining a fine-tuning pipeline or vector database.

A main-video approval writes an EDITED row per changed field plus an ACCEPTED row; a re-run writes SKIPPED. Short approvals record edits only, since their metadata starts from the video's draft. Feedback and the job's state change commit in one Firestore transaction, so the app can't remember an approval that never happened.

export function approvalFeedback(
  proposed: PickFields,
  final: PickFields,
  meta: { jobId: string; songTitle: string; pickVersion: number; at: number },
): Feedback[] {
  const rows: Feedback[] = [];
  const asText = (v: string | string[]) => (Array.isArray(v) ? v.join(', ') : v);
  for (const field of ['title', 'description', 'tags'] as const) {
    const before = asText(proposed[field]);
    const after = asText(final[field]);
    if (before !== after) rows.push({ kind: 'EDITED', field, before, after, ...meta });
  }
  rows.push({ kind: 'ACCEPTED', ...meta, after: final.title });
  return rows;
}

export function weighFeedback(feedback: Feedback[]) {
  return feedback.map((f) => ({
    weight: f.kind === 'SKIPPED' ? 'weak' : 'strong',
    kind: f.kind,
    song: f.songTitle,
    field: f.field,
    before: f.before?.slice(0, 300),
    after: f.after?.slice(0, 300),
  }));
}

Edits and approvals outweigh a re-run, since "try again" doesn't say what he disliked. Gemma gets the before/after pairs so it can see what he actually changed. Nathan can change his mind. He's allowed; it's his music.

Demo runs read his memory but write to their own jobs. Letting strangers teach the app what Nathan likes would be a spectacularly stupid feature.

How I Built It

I'm running Gemma 4 12B-it through llama.cpp's llama-server, with Google's open weights under Apache 2.0. The three “agents” are TypeScript prompt builders calling the same /v1/chat/completions endpoint with strict JSON schemas, which didn't need an agent framework. Gemma's 30-second audio limit sets the window size at 29.5 seconds, and the page advances one step per request because a 12B model on one L4 takes a few seconds to answer. I can work with that; pretending it's instant doesn't make it faster.

The page runs one pipeline step per request. A scaled-down, loading, or busy model returns a wait reason; reloading resumes the job.

Why Does Open Innovation Matter?

Nathan's unreleased music is the input. Gemma runs on a Vertex AI endpoint in my GCP project, so I don't send that audio to a separate inference API. Storage and uploads still touch GCP and YouTube, and Sentry gets traces with the media stripped, but I decide where the model runs and what reaches it.

Open weights also leave the door open to fine-tune on his music and audience data later, without replacing what I built this weekend.

Prize Categories

Best Use of Gemma 🪙

Gemma gets the audio and frames, not a description of them. chunk-analyst reports genre terms, instrumentation, vocal type, mood, and tempo feel in words. No lyrics.

Those terms drive a YouTube search of the genre's top videos. Hashtags and tags are picked from the job's candidate lists, with the artist name included in the tag pool, and tags must name something the analysis heard. The description's structure follows the audience evidence; its wording follows ARTIST_VOICE, distilled from Nathan's own YouTube descriptions: short, literal, his "hacked and slashed" credit line, the dry aside about what genre this even is. The title, description, and tags return a why.

Calls use a strict JSON schema, temperature 0.2, and a 2048-token budget. A bad parse gets one retry if the step has time left.

Audio numbers come from ffmpeg. The model is told not to produce them, and the server strips any it does. I don't need Gemma's opinion on what the meter probably said.

export function measuredFlags(m: Measurements | null): string[] {
  if (!m) return [];
  const flags: string[] = [];
  if (m.truePeakDbtp !== null && m.truePeakDbtp > -1) {
    flags.push(`True peak ${m.truePeakDbtp.toFixed(1)} dBTP, above the -1 dBTP ceiling`);
  }
  if (m.clippedSamples > 0) flags.push(`${m.clippedSamples} samples at full scale (clipping)`);
  for (const s of m.silences.filter((s) => s.end - s.start >= 2)) {
    flags.push(`Silence ${s.start.toFixed(1)}s to ${s.end.toFixed(1)}s`);
  }
  return flags;
}

Sentry Agent Tracing 🩻

Nathan is a fantastic musician and terrible with technology, so I'm also the support team. I already know the bug report: "it's broken," some unspecified time after it broke.

One Sentry trace per job: agent calls, duration, tokens, messages. Each invoke_agent span wraps the model's chat span. Audio and images become {type, mime, bytes} placeholders first; his unreleased track doesn't belong in my error tracker.

Two things the traces caught before Nathan did. First, every invoke_agent span was empty: Sentry only shows an agent's input and output from its own span, and I'd put them on the chat span underneath. Second, smart-pick was spending about 4,200 of its 8,192-token context on the prompt alone—each of his five previous uploads cost a description excerpt and a thumbnail. Three uploads still carry his credit and contact lines. A live pick dropped to 3,119 prompt tokens.

GitHub Copilot 🪛

Every PR gets a Copilot review before I merge it. GitHub Actions runs format, lint, typecheck, 1,050+ unit tests with coverage, and Playwright E2E on code PRs. CodeQL runs on code PRs, code pushes to main, and a weekly schedule; release-please handles versioning and changelogs. For a one-person build that has to outlive the weekend, the bot reviewer and the pipeline are the team.

Now He Has to Use It 🎬

Social posts and analytics can come later. First Nathan has to ship a real release with it and tell me what still annoys him. If I've done my job, the YouTube description stops being the unfinished task next to his music.

He's already plotting out the next set of features for me to build. He also says this version of AI is a lot easier to manage than the way he was doing it before.

Ashley ChildressFollow

Distributed backend specialist. Perfectly happy playing second fiddle—it means I get to chase fun ideas, dodge meetings, and break things no one told me to touch, all without anyone questioning it. 😇

🛡️ Apparently I Also Need an Editor

Claude helped draft this post, and Codex helped remove the parts that sounded like Claude helped draft this post—including writing this footer. I'm responsible for the claims and the app; Nathan is responsible for texting me when it breaks.