I'm a systems-first AI engineer. Instead of focusing only on models, I focus on what makes them run efficiently in production: distributed GPU orchestration, KV-cache routing, compiler-level cost prediction, and low-level Rust performance.
- 🎓 Bachelor's degree in Artificial Intelligence, currently completing a Master's in Data Science & Artificial Intelligence
- 🦀 Rust systems builder — from neural networks written from scratch to sub-100µs trading engines, with zero unnecessary Python runtime dependency
- ⚙️ Currently MLOps Engineer at Adorsys
- 🧠 Building compiler-grade tooling for AI infrastructure: intermediate representations, model-format interoperability, and inference-cost analysis
- 🔎 Open to MLOps / Platform / Rust backend engineering roles and infrastructure consulting
| Domain | Stack |
|---|---|
| MLOps / Infra | Kubernetes · Helm · vLLM · LMCache · llm-d · KEDA · Karpenter · Prometheus · Grafana |
| Rust Backend | Tokio · Actix-web · kube-rs · candle · serde · Rayon |
| AI Compilation | MLIR · LLVM · CUDA · Triton · ONNX · custom IR design |
| ML Engineering | PyTorch · TensorFlow · Burn · MuJoCo/PPO · CNN/LSTM |
Selected from github.com/rustnew for technical depth.
| Project | Stack | What it is |
|---|---|---|
| Lift | Rust · Custom SSA IR | A compiler framework built around one unified intermediate representation for both classical AI and quantum workloads — 13 optimization passes, O0–O3 pipelines, and LLVM/ONNX/QASM backends |
| UMC | Rust · GGUF/ONNX/SafeTensors | Universal Model Converter — moves AI model weights between formats (GGUF, ONNX, SafeTensors, and others) without silent quality loss, with the conversion's correctness kept verifiable rather than assumed |
| Rust_network | Rust · from scratch | A deep neural network implemented from first principles — forward pass, backpropagation, and training loop, no ML framework underneath |
| precog-trainability | Python · zero-cost NAS proxies | Research on whether signals computed on an untrained network can predict which hyperparameters will train well — reported alongside a headline result the project's own larger sample walked back, and a proxy that provably can't recommend one of its candidate methods |
Open to Rust / MLOps / AI infrastructure roles — let's talk.