In this interview, Paul Iusztin, founder of Decoding AI, walks through a series of pragmatic decisions made while building AI‑driven products. He explains why his team stopped using Retrieval‑Augmented Generation (RAG) for many scenarios, citing the 64k token context‑window limit and the latency‑inducing “zig‑zag” agentic retrieval loop. He critiques the industry‑wide hype around Multi‑Modal Communication Plugins (MCP), describing how a blanket adoption forced his team to write custom MCP servers that proved rigid and under‑utilised. Iusztin then shares his bullish view of “vibe coding” tools, arguing that AI‑generated code may render traditional clean‑code conventions less relevant, and predicts a future where AI evaluation (AI‑evals) becomes a critical discipline to ensure generated code works reliably. Finally, he discusses his personal choice to base his startup in Europe, highlighting lifestyle, tax, and remote‑work advantages, while acknowledging the trade‑offs of missing the Silicon Valley ecosystem.
Chapters
In our particular use case, everyone was using MCP everywhere. If you have hammer, you see nails everywhere. Load all the data into the context window, it will just sum up to 64,000 tokens at maximum. The model, if it doesn't find the answer, it will start to query the memory again and again. Zigzag adds extra latency. If I use Cloud Code and I just want to hook to my Gmail, I still think it's amazing. But when you build an application where you have full control over, I would just skip it. Usually answer was more in simplicity.
What went wrong with RAG in that application? You said you killed RAG.
We killed it, not entirely, but for many scenarios, we ended up not using it. For this particular use case, if we load all the data that an advisor has into the context window, it will just sum up like 64,000 tokens at maximum. In these scenarios, it's just easier just to load everything into a context window because when you do RAG, you have this retrieval problem, which is actually pretty complicated. And the model, if it doesn't find the answer, it will start to query the memory again and again and again until it finds. So it's basically this agentic loop, which is like a zigzag, which adds extra latency.
If I ask my intern, get this file and then double check if it's actually the stuff that I want, the data I want, most likely I think the human will go back and double check, right?
Yeah, that's a great analogy.
Yeah, because we, because they want to be helpful and they, when they, when they double check and they go back and then do it again, which that recursive loop could happen, which increase your cost and latency. When does it make sense?
When you have medium to big data or when you want to be super constrained on your context window. So basically the ratio between your data and your context window. I think that's like a good formula that you can find, do I need rag or I don't need rag?
You also said you regretted using MCP as well. While everybody is so excited about using MCP last year, what happened?
Everyone was using MCP everywhere. If one solution is hot, you shouldn't put it everywhere. Like if you have a hammer, You see nails everywhere. I see MCP as plugins very intuitively. And you can use these plugins in dev tools like Cursor, PlotCode. But we were actually building a product and we needed integrations with other providers like, like email. We thought every provider provides an MCP server and why shouldn't we connect to them through these MCP servers? But The thing is that our MCP server was sitting on our server and it wasn't hooked like directly to the client, like to the user. It was connected on our server. And ultimately we just ended up like writing custom MCP servers for each of our clients for all this integration. And we haven't used any of them because they were too rigid to our custom use case or they just weren't updated enough or they they just did not have yet an MCP server. If I use Cloud Code and I just want to hook to my Gmail to do some basic stuff, I still think it's amazing. Like, but when you build an application where you have full control over, then I don't know, I will just skip it.
What do you think of the, the, um, the very hot, like, vibe coding tools out there?
I'm pretty pretty bullish. Like I don't know how to code frontend. And even for me, a engineer, they can just extend my skills. Many people complain about vibe coding that it creates like messy code or code that doesn't respect like clean code or good architectural designs and things like that. But the thing is that these designs were made for people to avoid human error and maybe the AI will not need them anymore. Now we write Python that's compiled many, many times until it reaches bytecode. Do we read the bytecode to make sure it's clean bytecode? No, we just trust it.
And that's a good point.
We can see AI-generated code like just a compilation step.
We have to accept that the younger generation is speaking a different language. And then we just work with it.
Yeah, that's a great way to put it. We need like a way to ensure at the next level that everything works fine. AI evals, which basically ensures from a data point of view that your product works fine. And me personally, I think AI evals will be a huge domain in the future. Because if AI can code, then the only thing you can do is to evaluate it.
Yeah. Your startup is in SF. Why aren't you coming here?
I like Europe. I like the lifestyle from Europe. So that's one big choice. It's also very hard to move continents. In America, you need a visa, which makes everything more complex.
Yeah.
And I like my remote work setup. If I'm not constrained to be there, personally, I like more like to work by the sea, by the ocean, like in a more like, in a more relaxed environment. I don't like big cities.
Okay.
And agitation. Portugal is very connected to everything. It's closer to the US, to Europe, so I can very easily run a remote business. From there. In Romania, it's pretty hard because it's the most eastern part of Europe. And in Portugal, it's cheaper to live, lower taxes.
Do you feel you miss out on not being in Silicon Valley, or does it really matter?
Yeah, I do feel that I miss out, and it's okay. I chose this. It's a trade-off. It's a bet. You know, you cannot have it all.
