Last week, for the first time in four months, I didn't hit send.

I have written to you every single week for the last 16 weeks. It is a discipline I value deeply. But last week, I was underwater - descending, layer by layer, into a rabbit hole that started with a single piece of feedback and ended with the first white paper I have ever written.

It started when we built our first Hindi proof-of-concept for Spark. Technically, it worked. There was a moment of delight just seeing Devanagari script show up in Spark's interface.

But something was off.

I showed it to a friend. Her feedback was blunt: "This sounds like a government circular, Vaidy"

That one sentence sent me on a 2-week-long deep dive. The result is a 13,000-word report titled: Spoken by billions, misheard by AI.

Writing the paper gave me an education I wasn't expecting - in linguistics.

I came out of it with a new vocabulary. Some of it felt like rediscovering something I had always known but never had a name for.

"Diglossia" - the gap between a language's formal written form and how people actually speak it - is something every Indian has lived their whole life. We just didn't know there was a 9-letter academic term for it.

And then there was "agglutination." Which sounds like something that happens when you have too much gluten. But it turns out it is a linguistic term for how certain languages - most acutely the Dravidian ones: Tamil, Telugu, Kannada, Malayalam - build an entire sentence's worth of meaning into a single word by chaining suffixes together.

It exists elsewhere too, but nowhere does it combine with such thin training data and such a sharp spoken-written divide simultaneously.

I came away from this research with increased appreciation for these languages - and an awareness of how unfair the AI training data situation really is for them.

The research surfaced three problems - and I ended up organizing the paper around them as three quality gates every language Spark supports has to pass.

Does the AI hear it correctly? Hindi accounts for just 0.22% of all web data AI trains on. Tamil is 0.049%. English is 41.7%. That canyon of missing data is why AI mishears Tamil 6x more often than English. It's even worse for some other Indic languages..one widely used model hit a 125% error rate on Odia, hallucinating more words than it actually heard.

Does the output sound natural? Even with a clean transcript, AI writes Indic languages like a bureaucrat. Tamil has two completely distinct varieties - the formal written form and the version people actually speak. The gap is not "formal vs. casual." It is more like Shakespearean English vs. how you talk to a friend. AI defaults to the Shakespearean version every time.

Does it sound like this instructor? Two Hindi teachers - one who mixes 40% English into their sentences, one who does not - get nearly identical output from Spark today. The AI flattens what makes each voice theirs. That is the next frontier for us.

Mapping all three gates - and being candid about where we are on each - is what the paper is really about. And putting all of that out in the open is also what made me pause.

On one hand, I wanted to give back. The AI train is moving fast - but for Indic languages, it is moving fast while leaving a billion people behind..on the platform.

Anyone building for that billion is navigating the same gaps we ran into, and I wanted our research to save them some of the time we lost.

On the other hand - we are an early-stage startup. Is sharing all of this detrimental to our "moat"?

I've decided to lean into the "Open" theme of this journey. Trust is a better long-term moat than secrecy. And besides, this paper has already paid for itself internally. It isn't just a thinking tool - it has become a foundational knowledge base guiding every Indic language decision we make at Spark.

Whether you are a builder, or simply an Indian language speaker who has always wondered why AI sometimes sounds "off" in your mother tongue - I hope the research helps you as much as it has helped us.

Read the full white paper here

Onward!