AI / DevOps Engineer — five years across HFT, fintech and banking, building the full stack behind production LLM systems.

Currently at an international HFT hedge fund: self-hosted inference on vLLM and llama.cpp for confidential data, agentic and RAG pipelines, and the Kubernetes, observability and CI/CD platforms they depend on. I measure everything and optimise for cost per token, latency and reliability.

AI & Agentic Systems

• Self-hosted LLM inference on vLLM and llama.cpp with open-weight models (Gemma 4 E4B / 12B, FP8 and 8-bit), plus whisper.cpp for speech-to-text
• Inference tuning: quantization, speculative decoding, prefix caching, scheduler sizing — on power-capped GPUs
• Agentic and RAG pipelines, harness engineering, and a memory engine for AI agents
• Hybrid retrieval over PostgreSQL: BM25, pgvector, graph search
• AI application development with LangGraph (Python), rig (Rust), and eino (Go)
• LLM evaluation, tracing, and observability with LangChain, Langfuse, Prometheus, Grafana

Infrastructure & Platform

• Cloud and bare-metal automation with AWS, Terraform, Pulumi
  • pulumi-netbird
• Kubernetes (bare-metal clusters via Kubespray, Talos Linux, GitLab CI + Argo CD), including GPU node pools for inference
  • terradockernates, ansible-inventory-diagram, virt-cloud-init
• Monitoring: Zabbix, Prometheus, Grafana LGTM Stack, VictoriaMetrics, ELK
• Tracing & profiling: Grafana Alloy, Tempo, Pyroscope

Languages

• 🐍 Python — daily driver for automation, data pipelines, and AI applications
  • ansible-inventory-diagram, signup_api
• 🦦 Go — when simplicity matters
  • pulumi-netbird
• 🦀 Rust — when correctness and performance demands it
  • gitraider, nuttertools
• 🐚 Bash — glue that holds the universe together
  • dotfiles, virt-cloud-init
• 🐧 C — when feeling adventurous
  • gigachad_kernel_module

• The $250 server that replaces up to $7900 in API tokens — one month of self-hosted vLLM inference: 1 billion tokens for $250 versus $95–$7,900 on hosted APIs
• 612 output tokens a second on 70 watts: how to tune an AI server — prefill vs decode, speculative decoding, prefix caching and scheduler sizing, explained for non-specialists