Fifteen Years of the Same Problem

Someone always asks the same question when I show them what I am building.

Why would you want this? Let creativity stay human.

I understand the question. I have stopped treating it as an accusation, because the honest answer is not a defence of AI. It is just what I have been doing for fifteen years, one step at a time, and this is the step I am on.

Let me lay out the steps, mostly because I want to see them in order myself.

2011. I was CTO at Raintek Solutions in Mumbai and we built SMILE, an e-book platform for Indian regional languages. The thesis was that Marathi books were not reaching Marathi readers because distribution was broken, not because demand was missing. Bookshops were vanishing outside the big cities, print runs were shrinking, and the smartphone was arriving in every hand in Maharashtra. A phone is a bookshop that fits in a pocket.

We had commitments for a hundred books that were not in print. A hundred Marathi manuscripts with no other route to a reader. Some cheques had been signed.

Then my partner Shravan died of a heart attack. He was the editorial mind of the company and the reason writers trusted us at all. Raintek did not recover from losing him, and neither did SMILE.

For years I told myself the product was too early. It is a comfortable story and it is not the true one. Too early is what happens to products. What happened to SMILE was that a person died. I did not underestimate the technology. I underestimated how much publishing rests on people whose work is almost invisible, and how thin the whole structure is. An ecosystem like this can be knocked over by one heart attack.

The decade after. I did not stop working on Marathi content. I have spent about ten years running Bobhata, a Marathi infotainment publication, and that is where I learned the part nobody teaches you: what it actually takes to produce good work, consistently, at volume, forever.

The romance leaves quickly. Good writers have bad weeks. Deadlines arrive whether or not anyone has an idea. Quality is not a level you reach, it is a level you defend, every day, against tiredness and repetition and the simple fact that the twentieth piece this month is harder than the first. I have watched excellent writers file mediocre work because they were exhausted, and ordinary pieces come alive because someone had one good afternoon.

Ten years of that teaches you where a machine might genuinely help, and where it certainly cannot.

2025. So we ran the obvious experiment. For Diwali we produced SrujanMitra, which I believe was the first AI-generated Diwali Ank in Marathi. Not as a stunt. As a benchmark.

I wanted concrete answers to concrete questions. Does the machine have any feel for Marathi literature and culture? Can it carry emotion in the language? And which of these models is actually best at Marathi?

We ran the same brief through Gemini, ChatGPT, Claude, Grok and Deepseek, across essays, poetry and stories, and published every piece with its model named underneath. Human editing was deliberately minimal.

Two findings, one useful and one uncomfortable.

The useful one: Gemini was clearly the strongest at Marathi. Not close. If you are building anything in this language today, that is worth knowing, and it is the kind of thing you only learn by running the same brief through everything and reading the output properly.

The uncomfortable one: the poetry was bad. Genuinely poor. Marathi poetry carries an enormous amount in compression and restraint, and what came back was inflated and obvious. We published it anyway, with the model named, because the point of a benchmark is to show where the edge actually is. The stories held up better. Some of the essays were good. The poems were the clearest signal of how far there is to go.

We left the other errors in too. A story that turned solemn for no reason. Emotion laid on far too thick. Occasional slips in the language itself. Polishing those away would have defeated the purpose.

Now. Which brings me to long fiction, and why it is the next step rather than a swerve.

The Ank answered the short-form question. A poem holds together, badly. A short story mostly holds. Three hundred pages do not. Somewhere past the early chapters the machine loses the thread it was carrying, and everything I care about, continuity, memory, a character staying the same person, starts to come apart. That is the interesting failure, so that is where I went next.

I should be straightforward about what the system does, because it would be easy to hide behind the parts that sound harmless. It generates prose. It writes Marathi sentences, paragraphs and chapters. If a writer keeps what it produces, machine-written sentences end up in a finished book. The memory layer and the continuity checks are the part I find most valuable, but they are not the whole truth.

What two years of this has taught me splits along a line I first saw in Shravan.

He did two very different kinds of work. One was holding the manuscript: knowing a thread had been dropped in chapter three, remembering what a writer was reaching for two drafts ago, keeping the rules of a world straight so the writer did not have to. A machine carries that well. It never gets tired at the point in a novel where people usually stop.

The other was knowing whether the thing was alive. That has not moved an inch. What the machine writes unattended is technically correct and dead, and no prompt I have written has fixed it.

So the design follows the difference. The machine holds the memory. The writer keeps the judgement, sets the world, takes one of four directions or none, keeps or cuts every paragraph. I am not confused about who the author is.

Now, back to the question I get asked.

I notice it is asked most confidently in English. English writers can take these tools or refuse them, and many refuse, from inside an ecosystem that already has agents, editors, publishers, festivals and a global market. Marathi has a shrinking publishing economy, very little tooling, and a hundred books I know of that took years to reach anyone, if they arrived at all.

But that is the argument, and the argument is not really my reason.

My reason is that this is the thing I keep returning to. Fifteen years, four attempts, one language. Distribution, then publishing, then a benchmark, now the writing itself. I am not trying to replace anyone. I am trying to find out what is possible in Marathi with the tools of the moment, and I have been doing that since before these tools existed.

Shravan taught me that an editor’s first job was not to improve a manuscript. It was to know whether it was alive.

The machine can hold a manuscript together. It still cannot tell whether there is anything alive inside it. That part stays with the writer, and it is the part I am most careful not to touch.