LumiPost

I tested everything. E2E tests with Playwright on every feature, mocked Stripe and Supabase, green across the board. I felt good.

Then the first users signed up.

Within an hour I had DMs pointing to broken API calls. The backend was still hitting localhost:8000 in a dozen places. Not a logic bug, not a failed test. Just hardcoded URLs that never showed up locally because locally, localhost is correct.

Embarrassing? Yes. Surprising in hindsight? Not really. It's the kind of thing no test catches because it works perfectly in the environment where you wrote it.

The fix was trivial. The lesson wasn't.

Ship to a staging environment that mirrors production before your real launch. I didn't have one. I do now.

On the positive side: everything else held. The AI pipeline, the scoring, the auth flow, the Stripe integration... no surprises. Months of Playwright runs paid off.

LumiPost is live today. If you do competitive monitoring or industry research and spend too much time sorting noise, give it a try.

And if you find a hardcoded localhost somewhere — please DM me first. 🙂

https://www.producthunt.com/products/lumipost

When I started building, I had to make a few foundational choices fast. Here's what I picked and why.

FastAPI (Python) for the backend

The core of LumiPost is AI, vector embeddings, similarity scoring, summarization. Python has the best ecosystem for that, and FastAPI makes it easy to build a clean API without ceremony. No regrets.

Nuxt 3 for the frontend

I went with Vue over React mostly for the community. The Nuxt ecosystem felt more opinionated in a good way, less time debating architecture, more time building.

Supabase + pgvector

Postgres with vector search built in. Being able to store embeddings and run cosine similarity queries in the same database was a huge simplification.

The hardest part: I'm not a designer

Picking the stack was easy. Making the UI actually feel coherent and readable? That took way longer than I expected. No framework saves you from taste.

The choice I'd make again without hesitation: Mistral AI

Mistral lets you enforce structured JSON output natively. For a product where the AI pipeline is the core feature, that's not a nice-to-have. It's the difference between a reliable product and a fragile one. No prompt-hacking to parse responses, no edge cases blowing up in production.

If you're building anything AI-powered, that alone is worth the switch.

Last April, I was having coffee with a few freelancer friends. Different fields — marketing, consulting, finance. Same complaint: we were all spending 2+ hours every morning just catching up on industry news.

Not reading. Sorting.

Opening 20 tabs. Skimming headlines. Closing 18 of them. Repeat. By the time we actually found something worth reading, we'd already burned the best part of our morning.

We tried RSS readers. Too noisy. Google Alerts. Too generic. Paid tools. Too expensive for freelancers and for the result.

So I decided to build something that does the sorting for you.

The idea: you define your professional topics once, and the tool automatically scores every article against your profile using AI, so only the relevant stuff surfaces. No noise. No manual filtering.

I had no idea the hardest part wouldn't be scraping news or building the UI, it would be figuring out how to make “relevance” computable. That rabbit hole led me straight into vector embeddings and cosine similarity.

But here we are.

I'm launching LumiPost tomorrow on Product Hunt. Over the next few days I'll share how I built it, the tech choices, and what I learned along the way.

Follow along if you're curious 👇