ZeroClick

The zero-effort metadata pipeline: drop your video, get polished show notes and timestamps instantly.

Comment

ZeroClick Publish — Project Story & Submission Details

Elevator Pitch: Automate post-production admin with instant YouTube chapters, show notes, and click-ready metadata.

💡 Inspiration

Every creator knows the universal truth of content creation: the filming and editing is only half the job. The other half is the repetitive, mentally draining busywork that begins the moment you export your video. You find yourself scrubbing across an hour-long timeline just to mark down chapter timestamps, rewriting titles to appease the algorithm, digging through browser histories to find links to every tool you mentioned, and formatting descriptions for SEO.

This manual post-production admin routine steals 1 to 2 hours from every single upload cycle. It burns creators out and distracts them from what actually matters: storytelling and production quality.

We built ZeroClick Publish to eliminate that friction completely. The vision was simple: give creators a "drop-and-done" assistant where raw media goes in, and studio-ready chapters, show notes, and publishing metadata come out with zero manual scrubbing.

⚙️ What It Does

ZeroClick Publish turns raw recordings into ready-to-publish YouTube and podcast collateral within seconds:

  1. Automated Segment-Level Chaptering: Decodes audio, runs high-fidelity transcription, and semantically clusters topical shifts into platform-compliant YouTube timestamps (strictly respecting criteria: starts at 00:00, $\ge 3$ chapters, each duration $\Delta t \ge 10\text{s}$).
  2. Entity & Resource Mining: Listens for verbal mentions of software packages, GitHub repositories, books, creators, and gear, extracting and compiling them into a structured "Resources Mentioned" directory.
  3. Curiosity-Gap Title & Hook Generation: Analyzes transcript sentiment and key narrative peaks to propose three distinct title variations optimized for click-through rate (CTR) and character safety limits ($< 70$ characters).
  4. Interactive Scrubber & Instant Export: Provides an interactive studio dashboard where creators can audition clickable chapters against playback, tweak chapter labels inline, and copy clean, markdown-formatted descriptions directly into YouTube Studio with a single click.

🛠️ How We Built It

ZeroClick Publish is engineered around an asynchronous, local-first streaming architecture designed to handle heavy media workflows cleanly:

1. Audio Ingestion & Normalization

  • Uploaded media (.mp4, .m4a, .mp3, .wav) is parsed using FFmpeg to extract raw audio channels, resample to $16\text{ kHz}$ mono, and strip background DC offset and ambient hiss.

2. Time-Indexed Transcription Pipeline

  • Audio segments are transcribed via faster-whisper (CTranslate2) or Groq Whisper API for low-latency transcription.
  • Each spoken segment $S_i$ outputs exact start and end timestamps: $$S_i = \left(t_{\text{start}}^{(i)},\, t_{\text{end}}^{(i)},\, \text{text}^{(i)}\right)$$ which are mathematically normalized into standard $\text{HH:MM:SS}$ time codes.

3. Topic Clustering & Structured Extraction

  • Timestamped blocks are fed into Google Gemini 1.5 Pro / Flash using strict Pydantic / JSON Schema output constraints.
  • Mathematical constraints are enforced on chapter generation: $$\Delta t_k = t_{k+1} - t_k \ge 30\text{s}, \quad t_0 = 00:00, \quad K \ge 3$$
  • The model extracts topics, identifies named entities, and writes a retention-focused 3-sentence summary in a single zero-shot pass.

4. Reactive Studio UI

  • Built with Next.js, React, and Tailwind CSS.
  • Leverages the browser Web Audio API to link the playback timeline directly with the generated chapter cards. Clicking any chapter immediately jumps playback to that point.

🚧 Challenges We Faced

  1. Timestamp Drift & Hallucinations:
    • Problem: LLMs frequently invent timestamps that drift minutes away from the actual speech events when given raw text transcripts.
    • Solution: We anchored the LLM strictly to segment IDs generated by the Whisper engine. The LLM only chooses valid segment boundaries, guaranteeing zero hallucinated timecodes.
  2. Uploading Heavy Video Files:
    • Problem: Uploading multi-gigabyte 4K cuts over slow uplinks creates major bottlenecks.
    • Solution: Integrated client-side WebAssembly FFmpeg (@ffmpeg/ffmpeg) to strip and compress audio directly in the creator's browser before transmission, cutting upload payloads by up to $95\%$.
  3. YouTube Markdown Formatting Rigidity:
    • Problem: A missing 00:00 or malformed timestamp syntax silently breaks YouTube's chapter parser on the watch page.
    • Solution: Implemented automated regex validation passes on export to enforce deterministic platform compliance.

🧠 What We Learned

  • Schema Enforcement is Non-Negotiable: For creator workflows, unstructured outputs are useless. Relying on strict JSON schema ensures metadata can be piped directly into clipboards or APIs without breakage.
  • Speed Drives Adoption: In post-production, every second of processing delay feels amplified. Combining client-side audio stripping with high-throughput inference was critical to creating an instant-feeling tool.
  • Human-in-the-Loop Control: Automation should augment, not override. Providing inline editing before export gave creators the trust required to use the tool regularly.

🚀 What's Next for ZeroClick Publish

  • Direct YouTube Studio Integration: One-click OAuth publishing to update draft video metadata without opening the browser dashboard.
  • Social Short Cut-Points: Automatically highlighting the top 3 highest-energy segments with start/end marks for TikTok, Shorts, and Reels repurposing.
  • NLE Extensions: Packaging ZeroClick Publish as an Adobe Premiere Pro and DaVinci Resolve panel.

🏷️ Built With (Tags)

python fastapi gemini-api whisper next.js react tailwind-css ffmpeg typescript pydantic ai automation creators speech-to-text youtube-api hackathon

Built With

Share this project:

Updates

Submission history