I'm Avi. For a few years I've kept production automotive software running in C and C++, the connectivity side of the car: Bluetooth, WLAN, eCall, smartphone projection, and the deeply unglamorous defect-hunting that keeps them all talking to each other. Lately I've been climbing the stack into Rust and applied LLM/agent systems, mostly by building things in public and losing arguments with the Rust borrow checker. (I'm winning more of them now.)
Plenty of people can wire an LLM to an API. Fewer have spent years in the layer underneath, where "it mostly works" isn't a passing grade. I like living where those two worlds meet, and I'm looking for a team that does too.
A byte-pair-encoding tokenizer built from scratch in Rust. Deterministic, tested, and reassuringly fast. It's the first rung of a ladder that ends at a local tool-using agent, where every project is picked to force a Rust concept I've been avoiding.
A neural network that taught itself to play Snake, with no labeled data and no backprop — just a population of small networks, a fitness function, and generations of mutation and selection. Rung two of the Rust ladder, and a good excuse to actually understand evolutionary algorithms instead of just name-dropping them.
How I actually measure whether an LLM feature works, instead of squinting at it and hoping. Golden sets, precision@k, failure taxonomies, cost per task. The reusable spine every project below plugs into, and the part almost everyone quietly skips.
A working Model Context Protocol server that hands a real tool to an agent. The integration pattern every enterprise AI team suddenly wants, built and documented end to end, so you can see how the sausage connects.
A tool-using agent on one real workflow, with strict schemas, guardrails, memory, and an audit trail for when it inevitably does something surprising. Then run through the eval harness, so the claims are measured rather than asserted with confidence.
I think out loud about where AI engineering is actually heading: the harness over the model, the control layer over the demo, and the execution layer quietly moving to Rust. Opinions included, free of charge.
I'm a software engineer in Darmstadt, Germany, with an M.Sc. in Information & Communication Engineering from TU Darmstadt. For the last few years I've worked on production C/C++ connectivity software in the automotive multimedia world, hunting defects across Bluetooth, WLAN, eCall and smartphone projection with CAN traces, Wireshark, and an unreasonable number of ECU debug logs.
Outside of that, I spend my time building with RAG, LLMs, and agent systems, usually chasing the same question: how do you make AI actually behave inside a real, constrained, has-to-work engineering environment, rather than a demo that only cooperates on the third try? That's the throughline in everything on this site.
Lately I'm going deeper into Rust and agent systems, in public, one shipped project (and one humbling compiler error) at a time.
I'm looking for a team building real AI systems: applied AI, agents, edge, or the infrastructure holding it all up. If that's you, or you just want to argue about Rust or swap tips on getting LLMs to behave, my inbox is open (and refreshingly empty of spam).