Hobby projects
Things I build because I want them to exist.
It started with one archive for Sai Pallavi. It became a family of 15 archives, an algorithm platform you can watch, and an open-source tool for asking questions of your own files.
She turned down a fairness-cream endorsement, finished a medical degree while Premam made her famous, and keeps saying no to roles that exist only for glamour. Her films, from Fidaa to Gargi to Amaran, put ordinary women at the centre and let them decide.
I found her work by accident, watched everything, and wanted the films, the interviews and the ideas in one place, in six languages. So I built it: an independent archive, not affiliated, just admiration organised.
Sai Pallavi Archive
18 films with where to watch by region, 300+ interviews and performances found automatically from YouTube, and the ethics she speaks about. Tamil, Telugu, Malayalam, Kannada, Hindi and English.
How the first archive was built, in a day
The same discipline as my production work, applied to something I simply wanted to exist.
The brief became a locked spec before a line of code, the same gate every product of mine goes through.
Five translator agents worked in parallel while a fact-checker verified every trailer through YouTube’s own embed API and corrected dates and awards I had wrong.
Gemini validated the five translations against the English source and caught a real Tamil error; Codex and Grok reviewed the code and found a bug that would have silently dropped analytics.
Most readers are on phones. A Safari-only bug in the language switcher only showed up once the test suite ran in WebKit as well as Chromium.
Every Sunday a job re-searches YouTube and messages me only when there is something new. The page shows when its data last changed.
Watch links open each platform’s search rather than a hard-coded ID that rots; interview summaries are paraphrased, never quoted; the whole site says plainly that it is independent and not a way to reach her.
15 archives, one template
It started with one page for Sai Pallavi: her films, interviews and ideas in one place, in six languages. Then the same wish kept coming back for other people I learn from: where is everything they said, in order, without the clips?
So the page became a template, and the template became a family: podcast hosts, teachers, a spiritual guide, and homeschool lessons in course order. Each archive is independent and unofficial, plays on YouTube where the video was published, and refreshes itself every night.
The archives home
One directory for every archive: podcasts, learning, science, spiritual and cinema. Search it, open any archive, or request the next one and vote on what gets built.
What LeetCode doesn’t show you
LeetCode is great for practice. But a passing test tells you the answer worked, not why. I wanted to watch an algorithm think.
LeetCode gives a problem, an editor and a judge. WatchAlgo plays the algorithm step by step: every variable, pointer and data-structure change, with the code line that caused it.
Each solution is explained three ways: technical, fun (analogies and humour) and spiritual (patience, persistence, seeing structure). Same algorithm, three paths in.
Annotated code in JavaScript, Python and Java, plus the reasoning that leads to the approach — so the pattern sticks for the next problem.
How I built it: one developer, an agent pipeline
Before any agent ran, a content schema defined every field a visualization needs. The schema became the contract.
A single hand-perfected solution is fed to every generation as its pattern (retrieval-augmented generation). All other outputs inherit its quality.
12+ automatic checks on every output — schema, three languages, nine fields per step, three flavors. A failure goes back to the agent with the reason.
Sonnet with retries for bulk work, Opus for the golden example: routed by cost per successful output, not per call.
Drop a folder. Ask a question.
A personal RAG (retrieval-augmented generation: answers grounded in your own documents) that keeps your files on your machine.
RAG tools either upload your private files to someone else’s cloud, or need a vector database and an afternoon of setup.
Embeddings and chat run on your machine; zero external calls unless you choose a cloud model.
Everything lives in one SQLite file (sqlite-vec for vectors). No separate database to run.
One install command, under 90 seconds to the first answer from your own folder.
MIT-licensed, my first open-source project: built for other developers to run themselves.
Bring Anthropic, OpenAI or Gemini if you want; local models via Ollama by default.
Go and look
Start where it started, or jump to any of them. Each archive has a corrections link in its footer if you know of something missing.
The archives are independent and unofficial: not affiliated with Sai Pallavi or anyone they cover, and not a way to contact them.