Turn a pile of documents into a conversational, two-host audio overview — a NotebookLM-style "podcast generator" built on Backblaze B2. Create a show, drop in PDFs / articles / notes, and the app writes a two-speaker dialogue script with an LLM, renders it to multi-voice audio with TTS, and archives the sources + transcript + episode audio together under a per-show / per-episode prefix in Backblaze B2 cloud storage.

Each episode is a self-contained bundle of related objects — exactly the shape object storage is good at — which makes this a natural demo of B2 for AI artifact storage.

What you get out of the box:

• A Studio workflow: create a show, add source docs, generate a 2-host episode
• AI script generation (OpenAI) turning your sources into a Host A / Host B dialogue
• Multi-voice audio synthesis (OpenAI TTS), one voice per host, assembled into one MP3
• A Shows library scoped to the shows/ prefix — listen inline, read the transcript, download any artifact
• A podcast dashboard: shows, episodes, listening minutes, storage, generation activity
• The kept whole-bucket Files explorer and drag-and-drop Upload surface
• A FastAPI backend with strict layered architecture and structural tests, and agent-optimized docs

Dashboard — shows, episodes, listening minutes, and storage used, with a 7-day generation-activity chart and a recent-episodes table.

Studio — the three-step workflow to pick or create a show, drop in source documents, and generate a two-host episode.

Shows — the podcast library scoped to the shows/ prefix in B2, each card showing source and episode counts.

The pipeline is decoupled from upload: source docs live in B2, and generation runs as a FastAPI background task that fetches sources back from B2, calls the LLM, synthesizes audio, and writes the result. Episode status (pending → generating → ready | failed) is persisted in the episode's B2 manifest, and the UI polls it with TanStack Query, showing the blaze generating loader while work is in flight.

shows/{show_id}/show.json                              # show manifest
shows/{show_id}/sources/{filename}                     # uploaded source docs
shows/{show_id}/episodes/{episode_id}/episode.json     # episode manifest (status, voices, timing)
shows/{show_id}/episodes/{episode_id}/transcript.json  # structured 2-host script
shows/{show_id}/episodes/{episode_id}/episode.mp3      # synthesized audio
uploads/{filename}                                     # generic /upload destination

B2 is the sole data store — there is no database. All access goes through the S3-compatible API via boto3, contained in repo/b2_client.py, with a custom user agent set on the single shared client. See ARCHITECTURE.md for the full layering and data flows.

You need: Node.js >= 20, pnpm >= 9, Python >= 3.11, a free Backblaze B2 account, and an OpenAI API key (powers both script generation and TTS).

1. Install dependencies

pnpm install

2. Set up the backend

cd services/api
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cd ../..

3. Add your credentials

cp .env.example .env

Open .env and fill in:

• Backblaze B2 — head to the B2 dashboard:
  1. Create a bucket. Paste the Bucket Unique Name → B2_BUCKET_NAME, the Endpoint → B2_ENDPOINT, and the bucket's region (e.g. us-west-004) → B2_REGION.
  2. Create an application key with Read and Write permission. Paste the keyID → B2_APPLICATION_KEY_ID and applicationKey → B2_APPLICATION_KEY (shown once).
• OpenAI — paste your key → OPENAI_API_KEY. The default models (gpt-4o-mini for the script, gpt-4o-mini-tts for audio) and host voices/names are all overridable in .env.

Walkthroughs: creating a bucket and creating app keys. OpenAI keys: platform.openai.com/api-keys.

4. Run it

pnpm dev

Frontend at localhost:3000, API at localhost:8000. Open Studio, create a show, add a PDF or text file, and click Generate episode.

pnpm dev runs pnpm doctor first — a preflight that catches the common setup gotchas (wrong Node/Python version, missing venv, missing or placeholder .env, ports in use, missing OpenAI key) and tells you how to fix each one. Run it standalone with pnpm doctor.

• Show Creation & Source Ingestion — create a show, drag-drop sources into shows/{id}/sources/
• Episode Generation — LLM script → multi-voice TTS → archived MP3, run as a background task
• Shows Library — scoped explorer over shows/, inline player, transcript, downloads
• Dashboard — shows, episodes, listening minutes, storage, generation activity
• File Upload — generic drag-and-drop upload (also reused for show-scoped ingestion)
• File Browser — whole-bucket explorer over every object
• Metadata Extraction — image dimensions, EXIF, PDF info, checksums, audio fields
• Design System — tokens, primitives, the blaze generating loader, inline error/empty states. Live preview at /design.

• TypeScript, Next.js 16, React 19, Tailwind v4, shadcn/ui, Recharts
• TanStack Query — caching, dedup, retry, and status polling for every fetch
• Python 3.11+, FastAPI, boto3, Pydantic v2, PyPDF2, Pillow
• OpenAI — script generation (LLM) and audio synthesis (TTS), one SDK and one key
• Backblaze B2 (S3-compatible object storage) — the sole data store
• pnpm workspaces (monorepo)

MIT License - see LICENSE for details.

Manage Backblaze B2 from your terminal using natural language (list/search, audits, stale or large file detection, security checks, safe cleanup).

Repo: https://github.com/backblaze-b2-samples/claude-skill-b2-cloud-storage