Built by YGL · Knowledge & Productivity
InstaBrain
InstaBrain converts short-form video into structured knowledge. Send it a reel and it returns a titled, categorised, searchable note with the key points and what to actually do next — then makes your whole library answerable in natural language.
In active development · web and iOS builds
- Pipeline stages
- Fetch → transcribe → analyse → embed
- Search
- Semantic, over your own library
- Languages
- English · Hindi · Hinglish
- Capture
- Reels · screenshots · share sheet
The problem
People save hundreds of reels they never open again. The content genuinely holds value — a recipe, a tool, a tax detail, a place worth visiting — but it's locked inside a video: ephemeral, unstructured and impossible to search. The save button creates a graveyard, not a library.
The vision
Every reel you save becomes something useful. Not a summary you also never read — a structured note with the specifics extracted, filed where you'll find it, and answerable later by asking a question in plain language.
AI architecture
How the intelligence works.
Reel to retrievable knowledge
Several models collaborate in sequence. Each stage streams its status to the client, so a job that takes real time still feels responsive instead of frozen.
Fetch
yt-dlpAudio stream pulled from the source video. Session cookies from the browser extension unlock reels that require a login.
Transcribe
WhisperMultilingual speech-to-text producing a timestamped transcript with a detected language tag.
Analyse
GPT-4oStructured extraction into title, summary bullets, category, per-category metadata, searchable tags and a five-question quiz.
Embed
text-embedding-3-smallTranscript and summary encoded into a dense vector and stored alongside the structured record.
Retrieve
pgvectorQuestions are embedded and matched by cosine similarity across the user's library, with full-text search as a silent fallback.
What we built
The shipped surface area.
Everything below exists in the product's codebase. Nothing on this list is planned, aspirational or coming soon.
Multi-stage analysis pipeline from video URL to structured, stored note
Audio transcription that handles English, Hindi and code-switched Hinglish
Structured extraction: title, bullet summary, category, tags and metadata
Category-specific metadata — ratings and director for films, location and cuisine for food, instruments and risk level for finance
Auto-generated multiple-choice quizzes for recall on what you saved
Semantic search across your whole library via vector embeddings
Natural-language 'ask your library' retrieval with a keyword fallback path
Screenshot capture analysed with OCR and vision, through the same pipeline
Automatic routing into smart folders by topic
Live progress streaming while a video is being processed
Chrome extension supplying session cookies so private reels can be fetched
iOS app with share-sheet capture, so saving happens inside Instagram
Phone OTP authentication and subscription billing
Product experience
What it feels like to use.
Capture without leaving the app
Share sheet on iOS, extension in the browser, or paste a link on the web. Capture is one tap because a capture flow that takes six is a capture flow nobody uses.
Notes that tell you what to do
Extraction targets specifics — the tool's name, the actual steps, the location, the numbers — rather than a paraphrase of the video. The design principle is that a summary alone is not worth the tap.
Ask your library a question
'Horror films with good cinematography' finds the right saves even when none of those words appear in the note, because retrieval is by meaning rather than keyword.
Built for how India actually talks
Hinglish is a first-class case, not an edge case. Code-switched audio is transcribed and analysed, and notes can be returned in natural Hinglish rather than flattened into formal English.
Technical depth
The decisions that don't show up in a screenshot.
These are the parts that separate a demo from a product — and the parts most worth talking about with a technical buyer.
Every fallback degrades quietly
Semantic search falls back to full-text if the vector service is unreachable. The user always gets results; they never get an error screen because a dependency blinked.
Streaming progress over polling
Job state is pushed to the client as it advances through fetch, transcribe, analyse and store, instead of the client asking repeatedly whether it's done yet.
A frontend that owns its backend calls
The Next.js app acts as a backend-for-frontend in front of the Python service, so credentials stay server-side and the client talks to one origin.
Cost-shaped model selection
Transcription and analysis are separate stages with separately swappable providers, so the expensive model is only used for the step that needs it.
Stack
What it's made of.
Product
- Next.js
- React
- TypeScript
- Tailwind CSS
- Radix UI
AI
- Whisper
- GPT-4o
- OpenAI embeddings
- pgvector
Services
- FastAPI
- PostgreSQL
- Supabase
- Redis
Clients
- iOS via Capacitor
- Chrome extension
Operations
- Vercel
- Render
- Sentry
- PostHog
- Playwright
Current state
Where the product is today.
Stated plainly, including what hasn't happened yet.
Web application and iOS build are both running against the live analysis pipeline.
Transcription and structured analysis work end to end, including Hinglish content.
Semantic library search is implemented with a full-text fallback path.
Live on the App Store — no adoption or revenue figures are claimed anywhere on this site.
Work with us
Need this kind of engineering?
The practices on this page are the ones we bring to client work.