Quesma was founded in October 2023 as RefactorDB Inc. to build a database gateway. Overall, we failed, though the technology attracted prominent users and was acquired by our biggest customer, Hydrolix. We pivoted and continue our startup journey.
This is the postmortem of the database gateway. We had the right ingredients: $2.5M raised in two weeks, high profile pilots, technical complexity solved, strong problem validation and great team. We fell short on key factors: incompatible co-founders, mistaking validation for urgency, and the brutal question: are you a feature or a product? 95% of startups fail, but we usually hear the success stories. This is my attempt to fill that survivorship bias gap:
The background and equal co-founder trap
After 10 years at Sumo Logic (from 20-person startup to IPO), I spent 9 months exploring startup ideas. After 80+ customer discovery calls, I found a pattern: database changes are hard, and a proxy layer could help. Many proxies already exist, though they target a narrow use case (e.g. PgBouncer) while we could capture a wider use case.
I was obsessed with having an equal co-founder – smart people said it was a must-have. Paul Graham listed a “single founder” as the number one reason why startups fail. Though most of my friends prefer a career at bleeding-edge tech companies, a few others had different founder-market fit or were committed to other opportunities. I spent months on that, eventually convincing my former partner from Sumo Logic to join my venture.
In retrospect, this was my biggest mistake. I “convinced” a friend to be an entrepreneur – someone who is successful at established companies but didn’t have the zero-to-one entrepreneurship gene. A fatal mistake that backfired. Startups are not a smaller version of big companies. We ended up splitting after a year. I spent months recruiting my co-founder, then more time splitting up. We got a clean cap table, but I could have moved faster alone. I believe the stigma against single founders is outdated and wrong.
Instead, we should embrace more models. I was able to find excellent founding employees with great ownership. Without a co-founder, I can offer more equity, and in the Polish ecosystem, many great engineers see this as a compelling option. They want a solid paycheck on day one, but love early-stage vibes.
Lesson #1: A strong founding team with significant equity is more valuable than forcing an equal co-founder relationship.
The honeymoon: $2.5M and a two-pizza team
Raising our initial $2.5M proved to be doable. Being an early employee of a billion-dollar outcome helped get initial attention, as did my former boss who angel invested in me. We had a contrarian market thesis in a big market with many prospects validating the need for our product. Having prospects willing to speak with VCs helped too. Organizing a two-week roadshow with 50+ VCs was intense but incredibly fun.
Recruiting the initial founding team of six people and setting up the company (from legal through leadgen to office lease) took two months. We started with my co-founder working daily in September, and by the Christmas Party, we had a team and money in the bank. We rebranded from RefactorDB to Quesma. We also narrowed our focus to deliver an MVP: a proxy translating ElasticSearch queries to ClickHouse SQL.
Lesson #2: A two-pizza team is an excellent size for the pre-product market fit stage.
Words vs. actions: When validation isn’t enough
Our customer development focused on finding more potential customers, though a challenging trend appeared. We found big companies with strong problem validation. They accepted our solution, but we weren’t sure how to turn their interest into a real commitment without product.
Lesson #3: Require a concrete action from the customer (like running a script) to validate their intent.
Lesson #4: Cost savings need to make an impact at the top-line; generating new revenue is a much easier sell.
Database proxies are hard – Vitess took eight years at Google before PlanetScale monetized it. I was delusional about timelines. Our ElasticSearch-to-ClickHouse MVP took 9 months instead of 5. Early pilots uncovered flaws we fixed through Herculean effort (the pancake project was my favorite).
Moreover, our initial market proved to be way more niche. Though ELK stack showed up in a lot of LinkedIn resumes, many of those companies had migrated to a different stack or planned to. The best early adopters were already gone – they had migrated on their own before we launched. Instead, we targeted mostly big established companies that need more validation (late majority or laggards by Crossing the Chasm definition).
Some disagreements with my co-founder started to show up. I wanted to do more outbound vs. he preferred focusing on conference booths. I was pushing him to be more hands-on and show more initiative, while he perceived me as aggressive and chaotic. We avoided conflict, leaving our differences unresolved. We should have had a hard conversation around that time.
Forever in pilot purgatory
In summer 2024 we were sending out an early adopter version. We noticed a huge drop from verbally interested people to those who actually tried running it. The volume of potential pilots was low, though the profile was high, including a Fortune 500 company, a few established billion-dollar companies as well as a major telco. I hoped we’d speed up once we launched.
In August 2024 we landed our first major client with a simple painkiller solution. Their data platform got new prospects who demanded Kibana compatibility. We were their only option for that.
In October 2024 we launched around the time of KubeCon conference.
Lesson #5: Skip conferences booths pre-PMF; they’re too slow and too broad to drive meaningful learnings.
Meanwhile, our co-founder relationship deteriorated. We disagreed on fundamental decisions. A combination of travel, hard work, hectic high-profile pilot (in air-gapped env), and personal issues wore me out too. The traction was not hitting our goals. A lot of promising signs, but forever in the pilot waiting game.
After many hard conversations, we agreed to split up on reasonable terms. We announced it on the first work day of January 2025.
Are we a feature or a product?
Remaining a lonely founder was hard, but the team stayed strong and nobody else left. However, our biggest potential deal (six-figure ARR) went south. I pushed hard and generated more leads, but discovered a similar trajectory. After analyzing lost deals, we reached the same conclusion. We were never their top 5 priority. Usually we ranked around 12th priority at companies that only complete their top 2-3 items each quarter. We badly lacked urgency. If you’ve tolerated legacy software for years, why act now?
Around 20 companies use us regularly, mostly for free, and growing slowly. A few database companies want to partner, but it means more work for us. The biggest challenge was that we wanted to be independent, but end users just wanted us to fill the gap in their existing data solution. Moreover, AI coding assistants are making database migrations faster and cheaper, undermining the assumption that migration complexity creates sustained demand for proxy solutions.
Lesson #6: Are you a feature or a product? Could you evolve from feature to product?
I told the board that we should pivot. My VCs, Tomasz Swieboda from Inovo and Christian Jepsen from Heartcore, were extremely helpful. Luckily, our technology helped database companies close new deals. It neutralized a key objection, enabling some deals.
Our biggest customer, Hydrolix, grew 8x in 2024 and raised $80M in Series C.
The good failure: Selling and pivoting
We ended up selling our IP as an asset sale to Hydrolix. I’m grateful to Marty Kagan, Hydrolix’s CEO. This bought us runway for the pivot without bridge funding.
Many founders would position this as an acquisition success story. Failure is a cursed word. Although I learnt a lot, I failed on my first attempt.
I got some things right, but more terribly wrong. That’s where you learn the most. Reinforcement Learning works best when you move from 30% to 70%. Both for natural neural networks, such as human brains, as well as ones in Large Language Models.
We are on the next mission. I love my team, investors, and startup problem-solving. Stay tuned.
Stay tuned for future posts and releases
