Bravos AI
Build RAG-powered AI chatbots for your website. No code.
A year ago I didn't know what an API was.
I studied business in college, then sold real estate on Spain's Costa del Sol. International clients, dozens of languages, properties worth millions. But I kept seeing the same problem: a potential buyer visits your website at 2am from Saudi Arabia, writes in Arabic, and nobody's there. That lead is gone by morning.
I tried running an automation agency to solve this kind of thing for businesses. It didn't scale — every client was a custom project.
So I decided to build the product myself. The problem was I couldn't code.
I learned Python, FastAPI, and PostgreSQL by doing — mostly by talking to AI models. Claude, ChatGPT, whatever worked. The first version was terrible. But it worked.
Today Bravos AI lets any business create an AI chatbot trained on their own data in minutes. No coding. What makes it different from most chatbot builders is that it can actually query structured data like a database — a visitor can ask "apartments under 400k with a pool in Marbella" and get real filtered results, not approximate RAG matches.
It supports 12+ languages with automatic detection, captures leads, and the whole thing is live with a single script tag.
I'm a solo founder, still shipping features every week. Would love feedback from this community.
5 Comments
Thank you so much, Kevin! It wasn't an easy transition and I spent sooo many hours in the process, but I think it was worth it. As you say, the problem is real, and the solutions in most cases are very expensive for a small/medium size business that just needs a good chatbot. About the structured data querying... that was a nightmare to get working. It's a matter of giving the LLM context about the file it needs to answer from, and also implementing many layers of AI with different functions (filter extraction, rephrasing, etc) giving them as little responsibility as possible. And it is always combined with the semantic search, as a fallback in case the filters miss something. I was really surprised when I learned big platforms don't do this... so you can't really upload your catalog and get accurate answers. Because RAG is not very accurate on spreadsheets with 1000+ rows.
That makes a lot of sense. Breaking it into smaller responsibilities instead of letting one LLM handle everything is a smart move. I’ve seen the same issue with RAG on large structured datasets — it starts to get fuzzy really fast.
I’m especially interested in how you’re handling the filter extraction and mapping it to actual queries. Are you generating SQL directly from the LLM, or using an intermediate layer to validate it?
I’ve been working a lot with FastAPI + LLM workflows recently, so this is exactly the kind of problem I enjoy digging into. If you’re open, I’d love to help test or improve parts of it — even small things like query accuracy or latency.
Feel free to reach out anytime: kevin.chisumdev@gmail.com
Yeah... in my experience RAG and structured datasets are not compatible. RAG is just too vague for it.
Well, first of all I have to say it is not perfect. But acceptably accurate! And yes, I have an LLM layer that is just in charge of generating SQL filters from the user's natural language query. And in order for it to not make up filters, I feed it with the spreadsheets columns that the chatbot's admin has uploaded, and some values in them as examples. If it is not enough, there is a Rephrasing layer as well.
Let's do something! You can sign up in my app with a free account and as soon as I see your email I will upgrade your account for free so you can fully test it! I enjoy feedback, even if it is bad, so don't worry!
Thanks, Fabio — I just created an account with kevin.chisumdev@gmail.com.
Really appreciate the offer. I’m looking forward to testing it properly, especially the structured data query flow, filter accuracy, and overall response quality. I’ll spend some time with it and share honest feedback after I dig in a bit.