Journal
Notes from the studio
Essays on craft, technology, and the ideas behind what we build.
Answers that cite their sources
Building a conversational guide that never invents — the architecture behind a corpus-grounded assistant, and why “I don’t know” is a feature.
Read
What we mean by “Software with a Soul”
A studio is shaped by the things it refuses to compromise on. Here is ours — and why a slogan became a way of working.
Read
One chat, every language
Most of the corpus exists only in English. Most of our users do not. Rather than translate everything in advance, we taught the chat to translate as it speaks — and cached it so it stays fast.
Read
Where the thirty seconds went
Our assistant took 30–40 seconds to answer. The fix was not where we expected — and the lesson is older than any of our code: never optimize blind.
Read
Stories you can watch
Ask about a moment in Śrīla Prabhupāda’s life and the answer is no longer only text — it is a short clip of a disciple, remembering how it was.
Read
A library that tends itself
Five thousand lectures, gathered into one place — and a library smart enough to download what is next, delete what you have heard, and always keep something ready offline.
Read
Two of everything
To keep a library of lectures reachable from a village on 3G and from a capital on gigabit alike — even if the borders close — we mirror the entire app. Twice the cost, by design.
Read
The citation a model cannot fake
The cheapest way to stop a model inventing an identifier is to never show it one. A small trick — replace IDs with integers — that beats a model thirty times the price.
Read
A layered defense against invention
There is no single prompt that stops a language model from making things up. Grounding is a stack of techniques, assembled by experiment — and some of ours we threw away.
Read
A guide to the lectures, in plain words
Instead of scrolling through hundreds of recordings, you ask a question — and the assistant hands back the exact passages, ready to play.
Read