slobodan_

Hi, Slobodan here. I run a PTO/leave-management SaaS (Vacation Tracker, bootstrapped, profitable) with my co-founder for almost 8 years now, and this started as our internal tool.

We monitor competitors' websites and changes from the very beginning. It's one of the most annoying jobs (similar to PTO tracking), but we are not that bad at turning annoying things into products.

With some help, we shipped a prototype almost a year ago and shared it with a few friends. The idea was simple: get a short email with competitor website changes that actually matter every Monday morning. And 6 months later, everyone with access to the tool actually read these emails every week, so we decided to convert it into a real product MVP. But we built it in a way that's API and MCP-first, because we don't want another dashboard we need to log in to. The goal is to create a product with just 2 founders and a bunch of AI agents (we have 22 employees in Vacation Tracker, we want to see how far a product can go without hiring).

What Competitor Tracker does: we crawl competitor pages, store full snapshots (HTML, MDs, and screenshots), diff them, then classify each change (pricing, features, positioning, etc.) and write a short explanation with before/after evidence. We already track 2,000+ SaaS companies across 137 categories. In addition to that, we try to pull an older snapshot so new users have some results after ~15 minutes, instead of waiting a few weeks.

Everything is built using Cloudflare. We use Workers, D1, R2, Browser Rendering, AI Gateway, etc. I spent about a decade doing AWS serverless and co-wrote a Manning book on it, and this is my first product built entirely on Cloudflare instead. Happy to go deep on that tradeoff.

Honest limitations: extraction isn't perfect. We caught markdown extraction mangling prices, so price detections are now verified against screenshots. Error pages used to create phantom "changes" until we made them a distinct state. Classification sometimes overstates a page edit as a product change. JS-heavy sites are still hit and miss. We run automated QA passes on random samples to prevent users from seeing incorrect results.

Another cool thing was building a website and the product in a film noir detective theme with various checks (LLMs + deterministic) for voice and rules.

If this tool is useful to anyone else, feel free to sign up on the link above. You'll get 25 credits (we call them coins) for free (1 coin = 1 competitor/month). And just ping me if you need more credits.

Feedback is more than welcome!

True, there's a risk.

But if API Gateway or Cognito fails, I'll probably have a bigger problem than a tooltip that does not work. However, the tooltip will be cached, so if you loaded it once, it'll work even if API Gateway fails.

The AI part is used to generate tooltips, not to display them, so if it fails, I'll not be able to create new tooltips, or users will temporarily be unable to interact with them (ask sub-questions).

Creating a Tooltip Web Component using Svelte, API Gateway, Amazon Cognito Identity Pool credentials, AWS SigV4, and AI

On Day 25, we connects the Create Tooltip flow by writing React code and an AppSync Lambda resolver with some help from CofounderGPT and Github Copilot.

Thanks for posting the article :)

The structure of the article was an experiment, and it’s probably too long.

VideoPuppet is excellent. I am using it to create videos for the Five Minutes Serverless Youtube channel, and so far, results are outstanding. I can create a video from the markdown file really fast.

It's up again. It's WP, and not serverless unfortunately haha