I’m Ayaan, an AI engineer based in Surat, India. The short answer to what I do: I build AI systems that actually work in production. Multi-LLM pipelines, AI voice agents for live calls, workflow automations, RAG chatbots over real business data. For clients in the US, Europe, and wherever else the work takes me. The real story takes a bit longer.
How it actually started
I didn’t come up through a CS program or an AI bootcamp. It started with a PHP internship — backend logic, REST APIs, learning how real engineering teams actually ship software. Then a frontend internship, which forced me to care about the interface and not just the code running behind it. Neither was glamorous. But both taught me the thing most AI tutorials completely skip: what it actually takes to build software that survives contact with real users.
Everything AI came after that foundation, entirely self-taught. The loop was simple and a little brutal — build something, watch it break the moment someone asked an unexpected question, figure out exactly why, fix it, repeat. That’s how I learned RAG that retrieves the right chunk and not just the most semantically similar one. That’s how I figured out LLM routing that cuts a model bill in half without hurting output quality. And that’s how I built voice AI pipelines that keep end-to-end latency under a second. No classroom. Just a lot of late nights in Surat.
Why Surat matters — and why it doesn’t
I’m not downplaying the Tier-2 city thing. I’m leading with it. A self-taught AI developer from Surat, India, with a working portfolio and clients in the US and Europe, is proof that where you’re from doesn’t determine what you can ship. Timezone gaps have never been the real obstacle people expect them to be. Delivering solid, production-grade AI systems on time — that’s what actually closes the gap.
What I actually build
Four main areas, and I’ve gone deep on each:
- Multi-LLM systems and LLM routing. Designing pipelines that route each task to the right model instead of defaulting everything to the most expensive option. This is where real infrastructure cost control lives. A well-designed routing layer can cut LLM spend significantly with zero loss in output quality — and that shows up directly on a billing dashboard.
- AI voice calling agents. End-to-end voice AI systems built for live phone calls. Low latency, graceful handling of real conversation patterns — interruptions, silence, questions that go completely off-script — and pipelines that don’t loop or crash when callers don’t follow the expected path. These are production-grade, not proof-of-concept demos.
- Workflow automations with n8n and Make. Connecting AI outputs to real business infrastructure — CRMs, databases, Slack, email systems, internal tools. The goal is always something that runs reliably in the background without needing babysitting. If a human is manually moving data between tools, there’s probably an automation waiting to replace that.
- RAG systems and AI chatbots over real data.Retrieval-augmented generation built so the retrieval actually works — across messy, real-world queries, not just clean keyword lookups. That means chunking strategies that fit the data, embeddings chosen for the use case, and retrieval that holds up when users ask questions in ten different ways.
The FastAPI backends that tie all of it together are part of the package too.
How I’d describe the approach
Fast like a hacker when speed is what the moment needs. Careful like an engineer where it actually counts. I ship AI at vibe speed — but without the vibe-code mess that falls apart the moment a real user touches it. That contradiction is the whole point, and so far it’s held up under real billing dashboards and real client feedback.
If you’ve got an AI project that’s half-built and living in a Notion doc, that’s my favorite place to start. Let’s talk.