REELDEX

You saved it. Now find it.

A searchable second brain for Instagram Reels. Share a reel to @reeldex.io in a DM and it's transcribed, summarised, sorted by topic, and ready to search or ask questions about.

Live app · Features · Security · Architecture · Run locally

The problem

Everyone saves reels: a design tool, a recipe, an interview tip, a workout. Instagram's saved tab is a wall of thumbnails with no search, and you can't search what was said in a video. So finding "that free AI tool someone mentioned two months ago" means rewatching ten reels. Most of what we save never gets used.

How ReelDex fixes it

1. Share any reel to @reeldex.io in an Instagram DM, or paste the link on the web.
2. ReelDex pulls only the audio and transcribes every spoken word with timestamps.
3. AI pulls out what matters: a title, a summary, key points, the tools and links mentioned, promo codes, and step-by-step instructions. Then it files the reel under a topic. The creator's caption is used too: it fixes misheard names, and for music-only reels with text on screen it's often where the content is.
4. The bot replies in your DM with the summary and a link to the reel in your vault.
5. Search everything, or just ask: "What AI design tools have I saved?" returns an answer that cites the exact reels it came from.

Features

Your vault

Every saved reel in one place, with search across titles, creators, tools, and anything said in the video. Filter by topic, switch between a card grid and a table with a category sidebar, and watch new reels go from fetching to transcribing to done live.

Table view: category checkboxes and expandable rows	Select mode: bulk add to a collection or delete

Everything a reel said, on one page

Each reel has its own page:

• the summary, numbered key points, tools and links, and steps
• the full transcript with timestamps
• the creator's caption
• one-click translation to English for reels in other languages
• downloads as subtitles (.srt, with real timings) or .txt, or copy the transcript
• a button that opens the reel in the Instagram app on phones
• retry for reels that failed to process

Ask Dex: chat with your saved reels

Ask questions across your whole library in plain English.

• Grounded answers: every answer is built only from your own reels and links back to the videos and creators it used.
• Show more results: keeps digging past the first answer.
• Export: copy an answer formatted for WhatsApp, copy it as Markdown, or download a .md file.
• Saved chats: conversations save to your account, add new answers automatically, and can be reopened from any device.

Answers grouped by topic, each item linked to its reel	Source reels, WhatsApp / Markdown export, show more

Collections

Group reels into collections such as Design stack, Morning routine or Money moves. Each one gets a cover made from its first four thumbnails. You can fill a collection from your saved reels, rename it, or move single reels in and out.

Made for phones

Most reels are saved on a phone, so the app is built mobile-first. It has a bottom tab bar, full-screen reel pages and a thumb-friendly chat box.

     

Instagram, no password

Link your Instagram by sending a one-time code (MIND-123456) to @reeldex.io in a DM. The DM bot sends a magic link that signs you straight in. The token is removed from the address bar as soon as it's read.

Security and reliability

ReelDex handles other people's Instagram accounts and libraries, so access control is enforced on the server for every request. All of it is covered by 48 automated tests in tests/test_security.py, run by CI on every push:

• Ownership checks on every endpoint. Reels, collections and saved chats are scoped to the signed-in user. Guessing another reel's ID returns 404.
• Signed webhooks. Instagram webhook calls must carry a valid Meta X-Hub-Signature-256 HMAC, compared in constant time. The verify handshake uses a constant-time compare too.
• Exact-match pairing. A pairing code links an account only if it matches exactly and hasn't expired. Codes last 20 minutes and are single use.
• Bearer-token auth. Tokens travel in the Authorization header, never in URLs. Normal requests can't create accounts.
• Per-IP rate limits on sign-in, pairing, saving reels and AI endpoints. Over the limit, the API returns 429 with Retry-After.
• Input limits. Requests are validated against size limits. Thumbnail requests only accept real Instagram shortcodes and only fetch HTTPS images, capped at 5 MB.
• Locked-down server config. API keys can't be changed over HTTP, and the status endpoint reports only true/false flags.
• Duplicate-safe webhooks. Meta sometimes delivers a message twice, sometimes to two workers at once. Handled message ids are recorded in the database, so each reel is saved and answered once.
• Model fallbacks. Transcription, insights and Ask Dex each try a chain of models, so one busy or retired model doesn't take the feature down.

Performance

• Pay once per viral reel. If a reel was already transcribed for anyone, its transcript and insights are reused instantly with no AI calls. A lock stops two workers from processing the same reel at the same time.
• Audio only. yt-dlp pulls just the audio stream, never the full video.
• Lean list responses. The library list leaves out full transcripts, and collections and chat load their data in a single query instead of one per reel.
• Non-blocking AI. AI calls run on worker threads, so a slow model never stalls other requests.
• Instant, quiet UI:
  • the cached library renders immediately
  • polling every 3.5s while a reel processes and every 15s otherwise, paused in background tabs
  • no re-render when nothing has changed
  • lazy-loaded thumbnails
  • the Markdown renderer only loads when Ask Dex opens

Architecture

flowchart LR
    IG[Instagram DM] -->|Meta webhook<br/>HMAC signed| API
    W[Web app<br/>React 19 · Vite · Tailwind 4] -->|Bearer token| API[FastAPI]
    API --> DL[yt-dlp<br/>audio only]
    DL --> STT[Whisper large-v3-turbo<br/>timed transcript]
    STT --> LLM[LLM insights<br/>summary · tools · topic]
    LLM --> DB[(PostgreSQL)]
    API --> DB
    API -->|reply with summary + magic link| IG
    W -->|Ask Dex| RAG[Retrieval over your reels<br/>+ LLM answer with citations]
    RAG --> DB

Loading

Layer	Technology
Web	React 19, Vite, Tailwind CSS 4, lucide icons, sonner, react-markdown
API	FastAPI, Uvicorn, SQLAlchemy, Pydantic v2
Speech to text	Whisper large-v3-turbo on Groq, with OpenAI Whisper as a fallback
Insights & Ask Dex	Open-weight LLMs on Groq (GPT-OSS 120B/20B, Qwen) with a fallback chain
Media	yt-dlp, FFmpeg
Data	PostgreSQL in production, SQLite locally, with automatic column migrations
Messaging	Instagram Graph API webhooks and DMs
Hosting	Vercel (web), Render with Docker (API)
CI/CD	GitHub Actions on every push; auto-deploy to Vercel and Render

Running locally

Prerequisites: Python 3.10+, Node.js 18+, FFmpeg, and a Groq API key.

git clone https://github.com/imsayanpaul/Reeldex.git
cd Reeldex

1. API. Create a .env in the repo root:

GROQ_API_KEY=your_groq_key
# Optional: only needed for the Instagram DM bot
INSTAGRAM_PAGE_ACCESS_TOKEN=your_instagram_token
META_VERIFY_TOKEN=any_random_string
META_APP_SECRET=your_meta_app_secret

python -m venv venv
venv\Scripts\activate          # macOS/Linux: source venv/bin/activate
pip install -r requirements.txt
uvicorn backend.main:app --reload --port 8000

2. Web app

cd frontend
npm install
VITE_API_URL=http://localhost:8000 npm run dev

On Windows PowerShell, run $env:VITE_API_URL="http://localhost:8000"; npm run dev instead. Then open http://localhost:5173 and paste a reel link to try it.

3. Tests

pip install -r requirements-dev.txt
pytest

The suite uses a throwaway SQLite database and never calls Instagram or the AI providers.

Project structure

Reeldex/
├── backend/
│   ├── main.py            # App setup, CORS, startup migrations
│   ├── routes.py          # Reels, collections, Ask Dex, saved chats, webhook
│   ├── auth.py            # Bearer auth, ownership scoping, webhook signatures, rate limits
│   ├── models.py          # Users, reels, transcripts, collections, saved chats
│   ├── downloader.py      # yt-dlp audio extraction
│   ├── transcriber.py     # Whisper transcription with fallback
│   ├── summarizer.py      # Insights and topic classification
│   ├── search.py          # Ranking and Ask Dex retrieval
│   └── instagram_bot.py   # Webhook parsing and DM replies
├── frontend/src/
│   ├── components/        # Vault, reel cards & table, reel detail, collections, Ask Dex, dialogs
│   └── lib/               # API client, session, vault state hook, formatting & exports
├── tests/               # Security and access-control tests (pytest)
├── .github/workflows/   # CI: lint, build, startup on SQLite + Postgres, security tests
├── Dockerfile
└── render.yaml

License

Proprietary. All rights reserved. © 2026 Sayan Paul.

The code is public so it can be viewed and evaluated, for example by recruiters. It is not open source: copying, reusing, modifying, redistributing or deploying any part of ReelDex is not allowed without written permission. See LICENSE for the full terms.