I run LLMs in places they were never meant to run.
A Raspberry Pi in the middle of nowhere. Three serverless containers pretending to be one 27-billion-parameter brain. A language model from 2019, reanimated.
I'm a Senior Software Engineer in banking, building the data-quality and AI-governance platforms that regulators lose sleep over. In my free time, I build edge and distributed AI systems from serverless GPU clusters to UEFI firmware, usually to a hardcore bassline.
| Project | What it is |
|---|---|
| Commissure 🏆 | Distributed LLM runtime that splits Gemma-3-27B (~55 GB, unquantized) across three Cloud Run L4 GPU services, streaming bf16 activations over gRPC behind a single OpenAI compatible endpoint. Winner, Best of GPUs — Google Cloud Run Hackathon. |
| LLMiX 🏆 | "LLM-in-a-box" for Raspberry Pi 5: flash an SD card, boot, and get an offline AI Wi-Fi hotspot with chat, code help, Wikipedia RAG, and Medicanite, a fine-tuned medical model, top-scoring open medical LLM under 3B params. Built for the 2.6 billion people without reliable internet. Winner, AWS Breaking Barriers Virtual Hackathon. |
| KaiVM | The agentic hardware KVM: a Raspberry Pi that poses as a USB keyboard/mouse, watches the screen over HDMI capture, and lets Gemini 3 control any computer, even in BIOS or mid-crash. |
| ZombieLLM | Reanimation is all you need. GPT-2 XL (2019) brought back to life with knowledge distilled from GPT-OSS-20B (2025), an offline-first instruct model for laptops and Pis. On Hugging Face and Ollama. |
| RaveSQL | Lightweight Java framework that binds SQL files to methods with a single @SqlPath annotation — clean separation of SQL and code, automatic result mapping. |
- Languages: Java, Python, SQL, Bash
- LLM & ML: PyTorch, Hugging Face Transformers, fine-tuning & knowledge distillation, RAG, llama.cpp, vLLM, Ollama, CUDA, KV-cache internals
- Cloud & Infra: GCP (Certified Associate Cloud Engineer), AWS, Kubernetes, Docker, Terraform, Ansible
- Systems & Data: HPC grid computing, gRPC, Redis, UEFI / bare-metal, embedded Linux
Hardcore coding & speedcore basslines.