For seven years, I treated writing about building as a distraction from building.
It was a reasonable belief. Products do not ship because their founders post thoughtful threads. Customers do not stay because the CEO has a polished personal brand. When something was broken, I wanted to open the dashboard or the codebase, not LinkedIn.
So I built quietly.
More than ten products. Consumer apps, marketplaces, creator tools, AI products. Most failed. Two were acquired.
The result is a strange archive: years of decisions, numbers, mistakes, and reversals that almost nobody has seen.
I now think keeping it private was also a mistake.
The exits were not the whole story
One of the products we built was Hypee, a photo and video editor. It was sold to Yandex in a transaction valued at $3.1 million, and the technology became part of Zen.
That sentence sounds neat. Startup stories become neat after the ending.
Building them is not.
Another product, Imba, did not show enough promise as a standalone business. But its Android editor work later became useful because the Yandex integration needed Android as well as iOS.
Failure was not the opposite of success. In this case, a failed product produced a capability that mattered later.
Caramel, my first serious project as a full-time startup founder, taught a different lesson.
Caramel did not begin as an e-commerce marketplace. It began as a platform connecting Shopify stores with bloggers. Bloggers would promote sellers’ products and earn a commission. Shops liked the proposition: approximately 200,000 products were added in one week. Bloggers were less interested in working for commission.
So we pivoted Caramel into an e-commerce marketplace. We would aggregate the products and promote them ourselves.
Google Ads brought customers, but the unit economics did not work. We did not seriously test SEO or influencer marketing. Then we interpreted the paid-acquisition result as evidence against the product.
cheap.market later changed how I interpret that decision.
Its underlying model was surprisingly close to post-pivot Caramel: aggregate products, build a discovery experience, acquire demand, and earn a margin on transactions. The main sourcing difference was geography. cheap.market sourced from China; Caramel aggregated products from Shopify stores. Even the approximate margins were similar.
cheap.market showed that this model could produce real orders and workable early economics. Shopify’s Shop app is another market signal: it lets consumers discover products and stores from Shopify merchants through a personalized feed and purchase through the Shop ecosystem. At one point, it reached #1 in the Shopping category of the U.S. App Store.
Neither proves that Caramel would have succeeded. They changed my retrospective estimate, not the evidence we had at the time.
The mistake was more specific: Google Ads produced customers but not workable unit economics, and we treated that result as a verdict on the marketplace model. In reality, we had only tested one acquisition approach. We had not seriously tested the channels better suited to an e-commerce marketplace: SEO and influencer marketing.
With the marketplace and distribution experience we gained later, I think Caramel had a real chance to live.
This is one of the rare cases where I believe a lack of decent execution and experience killed the project - not the demand, the market, or the business model.
Speed was not the bottleneck
Years later, when Stable Diffusion appeared, I built the backend for an avatar product in approximately two days.
Then we delayed the release to add more features.
Lensa shipped a simpler product and captured the initial wave.
As an engineer, it is tempting to believe that the difficult part is building the system. Sometimes it is. But when the market is moving, the bottleneck may be deciding what not to build.
We later launched AI Boost. It reached approximately $40,000 in monthly revenue. I spent around eight months trying to push through its scaling ceiling.
The problem was not a lack of effort. The market itself was smaller than I wanted it to be.
That was harder to accept than an execution failure. Execution problems preserve hope: work harder, hire better, improve the funnel. A market ceiling asks a more uncomfortable question: what if you chose the wrong game?
This was where the accumulated disappointment caught up with me.
Hypee had been sold in a $3.1 million transaction. From the outside, that looked like proof that I understood the game. It did not translate into a comparable personal economic outcome, and the existing arrangement no longer looked like a path toward what I wanted to build.
Then we missed the initial avatar wave despite having the backend ready in two days. AI Boost reached real revenue, but eight months of work could not change the size of its market.
I had two exits on paper and a growing suspicion that I understood much less than I had believed after the first one.
The disappointment was not that the products had failed completely. It was that years of execution, a $3.1 million transaction, and a product at roughly $40,000 in monthly revenue had still not produced the entrepreneurial outcome I was working toward.
I spent exactly one week looking for a job.
That option felt even more depressing.
The week clarified one thing: I would rather live on crackers and in a pizza box than stop building companies.
That was not motivational rhetoric. It was the price I was willing to pay to keep doing the work I loved. The pizza-box line has aged less like a joke than I expected.
So I renegotiated the arrangement, decided to continue, and began studying companies while preparing AI Boost for sale.
AI Boost was eventually sold for $450,000. Once development and marketing costs were included, the sale was roughly at cost.
It was not a grand financial victory. It preserved the ability to continue and closed a chapter.
What changed
For roughly a year, I looked at private companies valued at more than $1 billion, startups that had recently raised capital, and the markets behind them. I wanted to understand the patterns behind market size, business models, distribution, and investor belief.
The order-of-magnitude differences were difficult to ignore. Depending on how the categories were defined, the niche available to a consumer editing product looked closer to hundreds of millions of dollars. A category such as CRM software was measured in tens of billions.
That does not make CRM an automatically good startup idea or every editor a bad one. TAM slides are easy to manipulate. But the comparison changed the question I asked. Winning a large share of a narrow category can produce less than winning a tiny share of a vast one.
Not because every company should raise venture capital. Not because valuation equals value. And definitely not because a unicorn label proves a healthy business.
I was trying to correct my own bias.
For years, I had treated the idea as a starting point and execution as the main variable. The research pushed me toward a different conclusion:
Market and business-model selection constrain what excellent execution can produce.
Execution still matters. It just cannot manufacture an enormous outcome inside a structurally small opportunity.
Why write now
The obvious question is: why did it take seven years?
Partly because I had the wrong model of writing. I thought the choice was between operating and performing expertise online. Given that choice, I would still choose operating.
There is also a strategic reason to start now: whatever I build next will probably need investment.
Investors usually see a compressed version of a founder through a deck, a few calls, and selected metrics. Publishing the actual decisions, mistakes, and reasoning creates a longer record. It can help the right investors discover how I think before a fundraising process begins - and help both sides decide whether we should work together.
That does not replace traction. A personal brand cannot rescue a weak company. But it can reduce the trust and context gap around a strong one.
But there is another kind of writing: preserving how a decision was made before hindsight cleans up the story.
A dashboard keeps the number. It rarely keeps the belief behind it. A git history shows what changed. It does not explain why the team believed that change would remove the constraint. An acquisition announcement records the outcome. It removes most of the wrong turns that made the outcome possible.
Once those details disappear, founders are left with slogans: move fast, listen to users, find product-market fit, choose a large market.
All true. Almost useless without the mechanism.
I want to preserve the mechanism: what we knew, what we assumed, which metric changed, what remained ambiguous, and which decision I would make with the same information today.
Writing does not replace operating. Done correctly, it makes operating experience transferable and makes the operator legible to future investors, partners, and team members.
The current chapter
The latest product was cheap.market, a cross-border marketplace.
It produced real operating evidence: approximately $8 paid CAC per buyer in live commerce experiments, a roughly $34 gross average order value, and measurable improvements in conversion to first purchase as we changed the product.
It also exposed how much work remained. Retention and engagement had barely received the same level of systematic attention. The company is now paused, and its future is unresolved.
I am not writing this from the summit.
I am between chapters.
The next objective is deliberately ambitious: build a company with the potential to become worth $1 billion.
I do not know if company number three will reach it. Anyone claiming certainty at this point is selling confidence, not evidence.
What I do have is an archive worth opening.
One commitment before the list: substance, not AI slop.
I use AI extensively, and I will keep using it for research, analysis, and editing. But every piece must contain something I actually built, measured, decided, or learned the expensive way. AI may help shape the material. It does not get to manufacture the experience.
If I have nothing real to deliver, I would rather publish nothing.
I will publish:
- company breakdowns focused on how markets were attacked, not just what products do;
- real operating numbers with definitions and caveats;
- product decisions that worked, failed, or produced the wrong lesson;
- applied AI systems that changed how we built and analyzed the business;
- the market and business-model filters behind the next company.
There will be successful decisions here. There will also be decisions that looked intelligent at the time and became obviously wrong later.
Those are usually more useful.
Experience that stays in one founder’s head is not an operating system. It is just expensive memory.
This newsletter is where I turn that memory into something reusable.
And this time, I will show the work while the ending is still unknown.
P.S. If you want the operating notes, numbers, and market breakdowns behind the next chapter, you can subscribe to Two Exits Later. The archive will remain on this site, and the emails will contain material I do not publish in the social versions.