Replica and DataPrysm, told as they happened: one company alive and running without me, one winding down with me.
I co-founded two startups. One of them is alive today, and I am not in it. The other is winding down, and I am. Told quickly, the first sounds like a success I walked away from and the second sounds like a failure I stayed for. Both summaries are close enough to be useful and wrong enough to bother me, so here are the longer versions.
Replica makes virtual try-on software: you see the thing on you before you buy the thing. It is a simple promise with an unreasonable amount of engineering behind it. I joined in August 2024 as cofounder and CTO. My whole time there lasted three months, and the three months were full.
The job was a rebuild. The product needed to be redone from scratch, and it turned out the org did too, so both happened at once, with clients waiting and everything in motion. It was my favorite kind of work: putting together a team and a product at the same time, with the hardest problem on the radar changing every day. By the end of my third month the new product was live inside client businesses and the company was at EUR 300K in annual recurring revenue.
Somewhere in the middle of all that, we interviewed at Y Combinator. A YC interview is ten minutes, which sounds like a strange amount of time to weigh a company, and is also, in fairness, more attention than most buyers will ever give you. Our ten minutes ended in a no. The partners did not believe in the idea. The clients, who had not been consulted, kept buying it.
In November I stepped back, for geography reasons. My cofounders kept going, and they run Replica to this day. Whatever dramatic ending the title of this piece promised, Replica declines to provide one. Companies are allowed to just keep working. Of everything in this essay, that is the fact I like most.
In April 2025 I went again, this time as cofounder and CEO. The title change matters to the story. At Replica the hard problems arrived in the codebase. At DataPrysm they arrived in the calendar.
DataPrysm was an agentic analytics platform. You connected your business data, asked questions in plain English, and got live dashboards back, with automated pipelines underneath keeping everything current while nobody watched. The first version did the unglamorous parts properly: OAuth into the tools a small business already ran on, a query layer that turned questions into something the data could actually answer, dashboards that stayed alive after the call ended. The timing looked right, too. In 2025 every business with a database wanted answers out of it without hiring a data team, and the models had just become good enough to turn that from a demo into a product. We called it “Lovable for Analytics”, a pitch that in 2025 needed no further explanation. We aimed it at small businesses, which I can now admit is another way of saying we aimed it at everyone.
The demo was good. I can say that plainly because the demo was never the problem. People watched it and liked it, over and over, in conversation after conversation. A good demo is a dangerous thing to own, because it pays you in enthusiasm on every call, and enthusiasm is the easiest currency in the world to collect.
We sold outward in rings. First the warm center: family, friends, our own networks, then the networks of our networks. When the warmth ran out, cold outreach. My commits got rarer and my calls got denser, which is roughly what the CEO title means at that stage. The volume was real, and for a while volume felt like progress.
Two things were wrong underneath, and both were quiet.
The first was targeting. We were talking to the wrong people. Not uninterested people, and not unkind ones, just not the people whose Tuesday our product would change, and we had never written down precisely who those people were.
The second was follow-up. There was no system that turned a good conversation into a next step. Notes lived wherever the conversation had happened, and almost none of them carried a next action with a date on it. Someone would like the demo, ask us to circle back in a few weeks, and the few weeks would arrive and leave without anything on our side noticing. No single missed circle-back feels like anything on the day it happens. The pipeline does not complain. Multiply that by months and you get a strange object: a pipeline that is full, warm, and completely still. DataPrysm never ran out of people who liked it. It ran out of follow-ups.
None of this was a shortage of effort. Effort was the one input we never missed. It could have worked. I say that without ceremony: the people were there, the interest was there, and the system that should have carried interest to revenue was not.
Along the way we applied to YC and did not get an interview. We sharpened exactly who the product was for, applied again, and that time we did. An interview date does something to a young company: for a few weeks the pipeline became a thing we prepared to describe rather than a thing we worked. The answer was still no. Two companies, two YC interviews, zero YCs. I mind that less than you might expect, and less than I did at the time.
The no came with homework: come back with traction. So the last months of 2025 went into getting some. We ran roughly 75 discovery calls and 35 demos, and six businesses started pilots on their own data. We signed a distribution partnership with a software company that wanted to put the product in front of its clients. Out of all of it came exactly one signed contract: GBP 300 for two licences, over three months. Readers keeping score will notice that the number 300 has appeared in this essay once before, wearing a K. That funnel, 75 conversations wide at the top and one contract wide at the bottom, is the two quiet problems from earlier written out as arithmetic. A second customer signed in the months that followed. At its peak, the whole company earned about 900 euros a month.
Then January 2026 arrived, and traction stopped being the main question. The big AI labs shipped DataPrysm’s core functionality inside their own products. Connect your data, ask in plain English, watch the dashboard assemble itself: the loop we had spent the back half of 2025 building became a built-in feature of the platforms we were building on. Everyone building on top of the models in 2025 carried the same quiet question, which was what happens when the labs ship this themselves. We carried it too. Knowing that weather exists and standing in it turn out to be different things. There is no meeting you can take about a model release.
So we pivoted. DataPrysm became a data and AI consultancy: MCP layers and data pipelines for legacy SaaS businesses that wanted what the labs were shipping and had systems older than their interns. MCP, if you have not met the term, is the plumbing that lets models talk to existing software, and a lot of software built before 2023 needs that plumbing custom-fitted. The first client was already in the building. The distribution partnership had outlived the product it was signed to distribute, and it turned into us being that partner’s consultancy across their 300 clients. The work was real, and a little of it still trickles in. But a consultancy that exists because a product could not is a different company wearing the same name, and we knew it. We are winding DataPrysm down now, carefully and on purpose.
Replica is still out there, run by my cofounders, selling software I have not touched since 2024. DataPrysm has a short list of things left to finish, and then it will be quiet.
When the two companies come up, people always want to hear about DataPrysm. It is the better story: a bigger swing, a louder ending, the big AI labs in the role of the weather. Replica, the one still alive and selling, gets a polite nod and no follow-up questions. I know exactly how that goes.