Architecture · scalable infrastructure · low latency
big tech HFT PSP big data fintech blockchain
16+ years of infrastructure engineering.
Security — confidential computing with TDX, SGX, AMD SEV and AWS Nitro, MPC, SOC 2
Low latency — Solarflare, kernel bypass, kernel patching, NIC firmware and IRQ tuning, Arista fabrics
Cloud and bare metal — AWS, GCP, Azure, OVH, and hardware I provision myself: custom KVM and libvirt layers, immutable images with mkosi and dm-verity, measured boot, Talos
Scale — past 10,000 nodes across regions, on cloud, on-prem and hybrid
Observability — Prometheus, Grafana, Loki, Tempo, Thanos, Vector, custom exporters
AI — I plan, build, verify, test, security-scan and maintain with AI tooling in the loop. I also build for AI: agents and agent tooling, MCP servers, LLM gateways, agent memory, and the inference infrastructure underneath
|
A Kubernetes controller that provisions GPU capacity from seven providers, selects the cheapest offer that meets the requirements, serves vLLM on it, and scales to zero when demand stops. GPU prices differ by multiples between providers. Karpenter supports AWS and Azure only, and Cluster Autoscaler requires node groups defined in advance. Neither can buy a spot GPU from a marketplace. |
Three Ansible playbooks deploy an LLM inference API on your own GPUs — batch endpoint, dashboard, per-GPU DCGM metrics, autoscaling, GitOps, WireGuard mesh across clouds. One variable selects Ray or gpuscale for capacity. Integrates K3s, KubeRay, vLLM, ArgoCD, DCGM, KEDA and WireGuard into one deployable platform. |
|
Cost, inventory and waste analysis across AWS, GCP, Cloudflare and OVH. Finds what is idle, oversized, previous-generation or unattached by combining inventory with utilisation, not one dimension. Runs as an MCP server and an HTTP API, so an AI agent can query spend and waste directly. |
Structural code intelligence for AI agents — tree-sitter across 40+ languages, trigram and inverted indexes, a dependency graph, behind 19 MCP tools. Answers structural questions — definitions, callers, dependencies, blast radius — in a few hundred tokens instead of whole files. |
|
A reference GitOps platform: cluster and services as reusable modules, Argo CD app-of-apps per environment, Rust microservices on Axum and sqlx, CI pipelines, and scripts that bring the whole thing up and tear it down again. Brings up dev, staging and prod as separate k3d clusters from one script, with CI, secrets and observability included. | |