Email deliverability startup
Coding
Product design
Brand design
Prototyping
Some learnigns
No code tools failed miserably, at least in this case.
Cursor and Claude are great helpers. They won’t spit out a working product just yet. But it saves a lot of time, on trivial things.
SQLite is amazing. Apparently one can have database, without running a database service.
Framer helped me ship landing page super quick. It also taught me bunch of Flexbox goodies (i.e. flex-order) or responsive layouts.
Framer vs. Figma. In Figma basic operations are super efficient. Few clicks, now questions asked if there is one right answer. It’s invisible, until you compare to less polished tool. And it begs interesting question... what level of interaction polish - is optimal im which context?
Context
A good friend of mine was consulting several companies on email deliverability. It works like this: client creates DMARC DNS record, and gets automated deliverability reports from Google, Microsoft, etc. But those reports are literally unreadable. So he approach me with a question... What if we try to build a startup - generate readable repots, help fix the problems. Let's really quickly test if this can become a viable business?
The story
I said "OK, I think I can hack us deliverability reporting” using n8n and LLM via API in two evenings”.
I spent first evening in huge frustration, tried bunch of no code tools, and went almost nowhere. I was stuck on extracting XML reports files from Gmail.
The next evening, I though, well... I’ve built integrations for Google apps before. So I took old good Javascript, and had a working prototype in one evening.
It took all unread emails, looked for relevant XMLs, extracted them, sent to OpenAI for interpretation, then sent interpreted report to the subscribers.
It worked!
But quickly we realized that results were a bit unpredictable - both, content and structure wise. Also it was quite expensive to run. So after trying to tweak prompts... I decided to try another way. Parse the reports using Javascript, store them in Database, and generate reports without LLMs altogether.
Then, if the report contained many errors, we would send it to LLM for actionable insights, and then full report would go to end-user.
It worked. We tested it on several new customers, and it really worked.
The irony is that it started with “No code + ChatGPT” and ended with "NodeJS + SQLite + Ubuntu + OpenAI API". LLMs actually helped to ship it much faster, but in a very unexpected way. Instead of LLMs doing everything, it helped me code faster. Cursor+Claude made me at least 2x more efficient coder.
OK. It works, what’s next?
Next is website. What’s the most effective way to put it live? I decided to give Framer a shot. Here is the result.
It remains to be seen if this become a viable business, but it was (and still is) amazing learning opportunity