Why I quit my PhD.

I quit my PhD after two months. Two months is barely enough time to develop a strong opinion about the cafeteria.

I don't think my department had fully processed that I had joined before I informed them I was leaving.

This was not the original plan. The original plan was very respectable. Master's. PhD or MBA. Become extremely qualified. Eventually start a company.

Instead, somewhere during my master's, I got irritated by experiments and started the company first.

The Beginning

I come from chemical manufacturing, which probably explains this

Other families discuss property and cricket at dinner. Mine discusses reactors.

I grew up around chemical manufacturing, which means I learned quite early that practically everything around us exists because somebody, somewhere, spent an unreasonable amount of time figuring out how to make it consistently.

The industry is incredibly sophisticated. It is also occasionally held together by one senior operator remembering what happened during a batch in 2007.

A surprising amount of industrial knowledge looks something like this:

"Why do we run it at 82 °C?"

"Because 85 doesn't work."

"Why?"

"We tried it."

"When?"

"Long back."

There is probably an Excel file explaining it. Nobody knows where it is. Or there are seventeen Excel files, one of which is called FINAL_v3_actualfinal_USE_THIS.xlsx. It is not the one you should use.

Then there is scale-up. A reaction works perfectly in a 500 mL flask. Everyone is happy. You put it in a 5,000 litre reactor and it develops a completely different personality. This is apparently normal.

And the usual scientific response is:

• Change something.
• Run it again.
• Change something else.
• Run it again.
• Ask the person who has been working there for thirty years.
• Run it again.

This bothered me. Not because trial and error is stupid. Trial and error is unavoidable. But because we were very bad at remembering the trials.

The Waiting Game

My master's thesis started because I got tired of waiting

Chemical engineering research involves a lot of waiting.

• You heat something. Wait.
• Mix something. Wait.
• Dry something. Wait.
• Send something for characterisation. Wait longer.

Receive the result. Discover that something went wrong twelve hours ago. Start again.

At some point during my master's, I realised I had become the slowest component in my own experimental loop. The machine could measure things faster than I could decide what to do next. So I started building software.

The basic idea was simple. Every time an experiment finishes, the system learns from it and suggests what experiment should come next. Then another layer checks whether that suggestion makes physical sense before someone spends three days discovering that it doesn't.

Basically: make the computer suffer through more bad ideas so the chemist doesn't have to.

I worked on it mostly at night. This was partly because I had actual master's work to do during the day and partly because almost every project I have ever cared about began as something I was technically supposed to stop working on.

Students should probably not take this as career advice. Although I did.

the computer gave me 40 materials

Eventually I let the system run properly. It proposed forty candidate materials. I had expected maybe five decent ones. Twenty-four survived the validation criteria we had defined.

That was the moment things became inconvenient. Because twenty-four out of forty is about 61%. The corresponding success rates I was comparing against were much lower.

So naturally, my first conclusion was not: "I have built something important." It was: "I have definitely coded this incorrectly."

I checked everything. Then checked it again. Then changed parts of the evaluation because surely there must be leakage somewhere. Then ran it again. Still there. Approximately 61%.

For context, I have never achieved 61% reliability in answering WhatsApp messages.

The system was also using substantially less computation because it improved the search as it learned.

That left me with two possibilities. Either I had made a very subtle mistake. Or there was something genuinely useful here. The second possibility was much more annoying because it meant I had to decide what to do with it.

Plan: Abandoned

My very sensible career plan lasted about another five minutes

Until then, the plan had been:

1. Finish master's.
2. Do an MBA.
3. Learn how companies work.
4. Eventually start one.

Nice. Orderly. There would presumably be certificates.

But now I had working software sitting on my laptop that appeared capable of doing something industrial R&D teams spend enormous amounts of money and time doing badly. "Eventually" suddenly felt like an unnecessarily long time. So I dropped the MBA idea.

My next extremely sensible decision was: Fine. I'll do a PhD.

The logic was actually reasonable.

• Spend five years going extremely deep.

• Validate everything.

• Build the science properly.

• Come out with a serious body of research.

• Then start the company.

This plan survived for approximately eight weeks.

A Failed Plan

My PhD career was not statistically significant

I started the PhD.

• Week one: very motivated doctoral student.
• Week three: doctoral student spending suspicious amounts of time talking to incubators.
• Week five: doctoral student who had somehow incorporated a company.
• Week seven: doctoral student who had built a demo and had not made comparable progress on the actual PhD.

This was not a particularly difficult pattern to detect.

By week eight, I had to admit something. I did not want to spend five years studying whether this could become useful. I wanted to find out whether someone would actually use it. Those are different experiments. And the second one had much faster feedback.

There was also a more uncomfortable question. What exactly was I staying for?

The answer, when I stripped away everything else, was partly credibility. A PhD would let me say: "Look, I am qualified to work on this."

Unfortunately, the software was already running. So I found myself in the odd position of pursuing a credential that would eventually give me permission to work on the thing I was already working on. That felt backwards.

The Decision

Then we started Novyte

Novyte started almost immediately afterward. In hindsight, taking absolutely no break between postgraduate research and starting a company was perhaps not evidence of excellent judgment. But it did make the transition efficient.

The master's project eventually became part of what we now call MatForge. The original problem was narrow: can a system learn from experimental results and choose better experiments?

The problem became much larger very quickly.

• What happens when years of failed and successful experiments are scattered across notebooks, ELNs, folders and Excel sheets?
• Can a system actually learn from an institution's experimental history?
• Can it understand when a process that works at 100 mL will stop working at 1,000 litres?
• Can it search enormous material spaces without asking researchers to test hundreds of hopeless candidates?
• Can the software become better every time the company using it runs another experiment?

That became the interesting problem. Not "AI for chemistry." There is already enough AI attached to nouns.

The goal became building a system that actually remembers how an R&D organisation works and helps it decide what to do next.

We eventually benchmarked MatForge across 22 different experimental problems and more than 132,000 simulated experiments. It outperformed the standard approaches we compared against on roughly 90% of those tasks. On average, it reached the target in substantially fewer experiments.

The remaining 10% are extremely useful because they prevent us from becoming unbearable.

We wrote the work up and submitted it for peer review. I now refresh publication portals with the same emotional stability people usually reserve for stock prices.

Unexpected Results

The much stranger part came later

The materials proposed computationally were actually synthesised. This was considerably more satisfying than watching another benchmark improve by 0.04.

There is a peculiar feeling when software proposes a material, somebody makes it in a laboratory, and the thing behaves roughly the way the system expected. Mostly because until that point everything is still numbers on a screen.

Then suddenly there is powder in a vial. Powder is very persuasive.

That was probably the moment the company became real to me. Not incorporation. Not the pitch deck. Not the website. Powder.

Reality Check

The powder then has to become a product

A material in a vial is a lovely thing. Commercially, it is also almost nothing.

Somebody now has to make ten kilos of it. Then a tonne. Then make the tonne behave exactly like the ten kilos. Then make it cheaply. Then make it repeatedly. Then make sure the reactor does not develop opinions.

This is the part of the industry I grew up watching, and it is where a lot of promising materials quietly die.

So that is where Novyte is pointed now. Not just: which material should we make? But: how should we make it? At what conditions? And what changes when we make ten thousand times more of it?

The same basic idea that chooses the next experiment can also help choose the next process condition. Temperature. Feed rate. Residence time. Solvent. Catalyst loading. All the knobs that turn a chemistry into a process.

And then comes the harder part: understanding what happens when you move from a flask you can hold in one hand to a reactor you could stand inside. Heat transfer changes. Mixing changes. Mass transfer changes. Things that appeared irrelevant at 500 mL suddenly become the entire problem.

The aim is that one day, "Why do we run it at 82 °C?" has an answer that is a model, a curve and a set of evidence. Not: "Sharma-ji said so."

I want to be careful here. Some of this works today. Some of it is still a roadmap.

Physics is complicated, reactors are stubborn, and anyone who tells you scale-up has been solved by software is probably about to send you a very expensive proposal.

We are not there. We are pointed there. And I think it is the right problem because it is the exact one I watched people lose sleep over when I was a kid.

The Real Lesson

I don't think the lesson is "quit your PhD"

That would be terrible advice. My sample size is one, and my PhD lasted two months. I am possibly the least qualified person to provide PhD-retention guidance.

The useful question, though, was this: what am I actually paying for? Not just in money. In time.

If the answer is deep expertise, access to equipment, extraordinary mentorship, a problem that genuinely requires years, or the life of a researcher that you actually want, fantastic. Stay.

But if the answer has quietly become: "I have already started doing the thing, but I feel like I need somebody to officially tell me I'm allowed to do it," then it is worth thinking about.

For me, the answer became obvious unusually quickly. Two months quickly. Possibly a department record. Nobody has confirmed it. I would prefer they don't investigate.

One final thing

I sometimes tell people Novyte started because I saw an enormous market opportunity in materials R&D. That sounds much more intelligent.

The truth is that it started because I got tired of waiting for experiments. Then I got tired of forgetting experiments. Then I got tired of watching companies repeat experiments somebody else in the same building had already run five years earlier.

Apparently irritation is a viable founder-market-fit signal.

We now spend our time trying to make materials R&D loops shorter: predict, make, test, learn, repeat. And trying to make the jump from flask to reactor boring, which is probably the highest compliment a chemical engineer can pay anything.

Not twenty years becoming twenty minutes. Physics remains annoyingly non-negotiable.

But twenty experiments becoming eight? Six months becoming six weeks? A process scaling the first time instead of the fourth? A team remembering everything it has ever learned? Those are very real possibilities.

Human civilisation likes naming ages after materials. Stone. Bronze. Iron. Silicon.

I'm not arrogant enough to name the next one. I would, however, like Novyte to help discover whatever it is made of. And then make sure it still works when someone orders five tonnes.