I build the AI infrastructure I'd want to use.

I'm Alberto Bermudez. Today I'm the founding AI engineer at Tekkr, where I ship production AI systems — RAG platforms, agents, internal tooling — that real teams use to fix operational bottlenecks.

Most AI engineers reach for an LLM first. I don't. I start with the problem and work backwards to the simplest system that solves it. Sometimes that's an agent. Sometimes a vector database. Sometimes a classical model that runs in milliseconds and costs almost nothing. The tool follows the problem.

That approach comes from spending time in very different environments before I landed in AI engineering full-time.

Siemens Energy (5 years, technical sales) taught me to sit with the customer before proposing a single line of code. The product is the conversation; the implementation is what falls out of it.

Siemens AG (eMobility) had me building data pipelines for thousands of EV chargers. The lesson there was unambiguous: most problems are data problems before they're model problems. Schema, latency, observability — that's where the work lives.

Siemens Mobility hosted my M.Sc. thesis at TU Berlin: a GraphRAG assistant for safety-critical codebases that reduced developer onboarding time by ~30%. Final grade: 1.3 (sehr gut). It's the work I'm most proud of — and the one I keep iterating on as a product.

The path that got me there: a B.Sc. in Electrical Engineering at TU Berlin gave me the grounding, then the M.Sc. let me specialise in ML and AI. EE first, then ML, then production — in that order. It's a longer route than most, but it's why I look for the simplest system, not the flashiest one.

On the side I work through the 42 Berlin curriculum: a custom shell, a raycasting engine, an HTTP/1.1 server, philosophers, push_swap, the full C++ module sequence. AI engineers rarely write a malloc'd line of C — and that's exactly the muscle I want to keep.

Berlin is home. I'm open to talking about production AI systems, GraphRAG, retrieval, and anything where the product question comes before the model question.