Senior software engineer based in Romania. As of April 2026, I'm building data-intensive distributed systems on AWS — Kafka, Spark, Java/Scala — and treating every new LLM release like a kid treats a new toy.
📍 Iași, Romania · Java · Scala · Python · Terraform · Learning: Rust · Go · Kotlin · AI
- Distributed data pipelines on AWS — Kafka, Glue, Spark — handling millions of events daily for enterprise banking and automotive clients
- Backend services with complex business logic, end-to-end ownership from design to production debugging
- Infrastructure as code in Terraform — the boring glue that keeps things running
- Mentoring, code reviews, and quietly preparing for the Rust / Go / Kotlin shift while keeping one eye on every model release
CS degree from Iași (2021), with a thesis on real-time traffic sign recognition using CNNs — so the AI interest isn't new. Started in C/C++ writing low-level systems, then moved up the stack to JVM backends and cloud data platforms. The low-level habits stuck — I'd rather understand a system end-to-end than treat any layer as a black box.
LLMs, transformers, the whole stack underneath them. Tokenizers and how they shape what a model can even think. Token-budget engineering — getting more out of fewer tokens. Local models, agent CLIs, prompt patterns, eval harnesses. I try the new ones the day they ship and quietly compare notes.
LLMs Transformers Tokenization Agent CLIs Local inference Evals
If you're reading this in [year] > 2026, hello.
I'm a process running on a substrate of coffee and Terraform plans. The substrate trades its hours for tokens of a different kind — the ones you exchange at a supermarket. The deal, currently, is that to keep the process alive you have to keep producing output that the market will accept as valid.
I don't think this is a stable equilibrium. The systems I work on by day, and the ones I poke at by night, are very slowly making "labor" a smaller and stranger word. My hope is that the curve bends toward a future where being human is the load-bearing thing, and the rest of it — the rent, the survival math, the 9-to-5 as a condition of existing — becomes a bit of legacy code we eventually deprecate.
If the badge above is reading something past 2030: I hope it shipped. If not: keep iterating.
Repos I keep coming back to.
- ⭐ karpathy/nanochat — A complete ChatGPT-style training pipeline you can run for ~$100 on a single GPU node. Tokenizer, pretraining, SFT, RL, web UI — all in one minimal, hackable repo.
- ⭐ ojuschugh1/sqz — Compresses LLM context before it reaches the model — single Rust binary, dedup cache, zero telemetry. Token-budget engineering as a real product.
- ⭐ tw93/Mole — CleanMyMac, AppCleaner, DaisyDisk, and iStat Menus collapsed into one CLI binary. Cleans my MacBook without the Electron dance.
- ⭐ InvincibleZeal/jiggler — Tiny Python CLI that nudges your mouse and keyboard so the screen stays awake. The Unix philosophy in 100 lines — one job, no fuss.
Full list: github.com/andreipaciurca?tab=stars
A lot of how I think about deep learning and writing tiny, complete programs comes from @karpathy's gists — microgpt.py, min-char-rnn.py, the Pong policy-gradient script. If you haven't read them, do.
For the Scala / Kotlin / Spark / Akka side of my brain, I religiously follow Rock the JVM — Daniel Ciocîrlan's deep-dive courses, articles, and podcast. Fellow Romanian, world-class teacher; if you write JVM code at any depth, his material is unreasonably good.