TubeNotes

An independent SaaS product that turns YouTube videos into AI-generated notes and summaries.

Stage
Publicly available
Role
Founder and builder

Kind of work

  • Product development
  • APIs
  • AI integration
  • Data pipelines

What it does

TubeNotes takes a YouTube video and returns structured notes and a summary that can be read, kept and revisited. It is a subscription product, with accounts, saved history and usage limits.

My role

I founded TubeNotes, built it and run it. I am responsible for the application, the data infrastructure, the analytics and the product decisions.

What was built

  • A web application, in TypeScript with Next.js, for accounts, the summariser, saved history and billing.
  • A Python API, built with FastAPI, that runs the pipeline turning a transcript into notes.
  • A PostgreSQL database holding users, summaries, usage and a record of every pipeline run.
  • Subscription billing with usage limits, through Stripe.
  • An analytics warehouse in BigQuery, modelled with dbt from the product’s own events, with data-quality tests for freshness, unique identifiers and lost events.

Engineering challenges and decisions

  • Long transcripts. A long transcript is too large to handle well in one request. The pipeline splits it along topic boundaries, extracts structured items from each part in parallel, then writes the notes from those items.
  • Missing captions. The service falls back between caption sources and languages. A video with no captions can be transcribed from its audio as a paid, opt-in background job. Credits are reserved first and refunded if the job fails.
  • Unreliable model output. A model call is retried a fixed number of times, and malformed JSON is repaired where possible. If one section of a long video fails, the remaining sections are still returned.
  • Knowing whether the output is good. Every run is recorded stage by stage, with timings, token counts and prompt versions. Rule-based checks and a second model score the output afterwards, for monitoring.

Outcomes and current status

TubeNotes is publicly available. Everything listed above is implemented in the product.

Its present limits:

  • All text generation uses one language-model provider.
  • Output scoring monitors quality. It does not yet stop a poor result from being returned.
  • Pipeline changes are compared on a fixed set of videos by a harness that is run by hand.

Technologies

  • Python
  • FastAPI
  • PostgreSQL
  • Next.js
  • TypeScript
  • OpenAI API
  • Stripe
  • BigQuery
  • dbt
  • Docker

Visit TubeNotes