Krishna Varma

Ruminations on software, running, sanskrit, etc. Essays at

[Krishna Varma @udbhava]

No words can express the grief of losing those you love most. Performing their last rites yourself, a ritual this raw where you're handed the fire, hits even harder. And yet it is the ritual that carries you. Grief needs a container, and ritual is that container. You fold your hands and chant the same mantras, the same sequence of rice, sesame, water and fire performed in millions of Hindu homes, unbroken for millennia.

Byung-Chul Han says rituals are to time what a home is to space: they make time habitable. Modern life is starved of ritual. We never lost it.

[prathosh A P @prathoshap]

Dream 2, realised - Introducing Vāgbodhini (वाग्बोधिनी)- a personalized tutor for Sanskrit.

1. Input any text (any script) and hear a flawless, metrically precise chant.

2. Chant it back and get instant, syllable-by-syllable phonetic feedback. 🧵

Link in the First Reply.

[Krishna Varma @udbhava]

@fewmorelines Update: this is fixed and live now, web + both apps. Thanks again for the catch. Also just shipped Tamil and Bengali scripts.

[Krishna Varma @udbhava]

While building the Gita app I was genuinely surprised how well the semantic search turned out. I typed "how do I stop worrying about how things turn out" and the top result was 2.47, on doing your work and letting go of the outcome. It reads the intent, not the phrasing.

[Krishna Varma @udbhava]

@fewmorelines Good catch, thank you. You're right, conventional Kannada uses the anusvāra here (ಸಂಜಯ, ಪರಂತಪ); we're emitting the Sanskrit-faithful conjunct (ಸಞ್ಜಯ). It’s systematic, so I'll try to fix it across the text.

[Krishna Varma @udbhava]

@chvidyasaga22 Not RAG, just semantic search. Finds the closest verses and shows them untouched, nothing generated.

[Krishna Varma @udbhava]

You can try it here: https://t.co/7w6eMVntGz

[Krishna Varma @udbhava]

Thanks for the close listen, this one is intentional. The app recites the verse pada by pada for the read-along, so each pada ends at a pause. At a pause the correct form is one with the visarga restored, पापेभ्यः, not the sandhi form पापेभ्यो, which only holds when the word runs straight into मोक्षयिष्यामि with no break. So पापेभ्यः is right here given the pada break.

[Krishna Varma @udbhava]

You open-sourced Vāgdhenu a week ago, and it's now the voice of my Bhagavad Gītā reader! Thank you @prathoshap 🙏

https://t.co/7w6eMVntGz

[prathosh A P @prathoshap]

A 15-year-old dream has come true today. I started a PhD with the dream of creating a system that chants any Sanskrit shloka perfectly.

And here I am opening sourcing 𝐕𝐚𝐠𝐝𝐡𝐞𝐧𝐮 - 𝐀 𝐯ṛ𝐭𝐭𝐚 (𝐦𝐞𝐭𝐞𝐫) 𝐚𝐰𝐚𝐫𝐞 ś𝐥𝐨𝐤𝐚-𝐭𝐨-𝐜𝐡𝐚𝐧𝐭 𝐭𝐞𝐱𝐭-𝐭𝐨-𝐬𝐩𝐞𝐞𝐜𝐡 (TTS) 𝐬𝐲𝐬𝐭𝐞𝐦 𝐟𝐨𝐫 𝐒𝐚𝐧𝐬𝐤𝐫𝐢𝐭. This is the world's first vrutta-aware, open-source TTS for Sanskrit Chanting.

[Krishna Varma @udbhava]

@anilsahoo___ @prathoshap Thank you 🙏 Fair point, the translation & Shankara's commentary can read dense. I’m biased towards keeping them faithful rather than paraphrasing, so the app never puts words in the Gita's mouth. Which part felt hardest, the translation or the commentary?

[Krishna Varma @udbhava]

@opdroid1234 Nice, the input side of the same problem. A solid Indic keyboard + voice input is a missing piece. Keep us posted!

[Krishna Varma @udbhava]

@prathoshap Exactly. Rendering pada by pada means each is spoken in its pausal form: visarga sandhi and jastvam too plus anusvara. Keeps every pada a clean standalone utterance.
On the M4 it ran ~25s per pada, so the full corpus came to roughly 18–20 hrs.

[Krishna Varma @udbhava]

Thank you for catching this. It's an automatic transliteration from the Sanskrit. It's likely tripping on specific conjuncts, and I really want to get it right. Could you DM me a verse or two showing what's displayed vs what it should be (a screenshot is perfect)? That'll help me fix it 🙏

[Krishna Varma @udbhava]

@prathoshap Own hardware, all local, not the demo. Ran the released code + weights on a Mac mini (M4, 16GB) via MPS, rendering pāda by pāda with a fixed seed, then stitching. The demo's daily cap made local the obvious route and the M4 handled the full ~700 verse corpus without trouble.

[Krishna Varma @udbhava]

@prathoshap Thank you for open-sourcing it 🙏 I render each verse a pāda at a time with it and stitch them, so the read-along can follow the chant pāda by pāda as it plays. The chant came through consistent enough to build on cleanly, which made the whole thing possible.

[Krishna Varma @udbhava]

@prathoshap Incredible work @prathoshap! Will try this out for my Gita reader app. Happy to contribute in any way if it helps the project.

[Krishna Varma @udbhava]

@arjuna_uvaacha And beautifully, the Gita echoes the same: Krishna says among sacrifices, He is the Japa yajna (BG 10.25). The Sahasranama's conclusion by Shiva himself is that very Japa personified. The tradition really wrote in circles of perfection 🙏

[Krishna Varma @udbhava]

This hits even harder now that AI writes the code. It's great at the happy path, tutorial-perfect code that compiles and looks shippable. But that's still just the 20%.

The other 80% is still on you: testing it under real load, verifying it isn't quietly wrong, debugging the failure modes it never saw coming. AI is industrializing the happy path. The harder 80%, the actual curriculum, is now the whole job.

[Krishna Varma @udbhava]

@hsraghav The vairagya that arrives like this, the moment impermanence is actually seen, is the kind that hits home the most. Not renounced by effort, but revealed by clear sight. As Raja Rao eloquently put it: "Truth began where sorrow was accepted."