02 / What I do

Experience

AI engineering on top, automation through the middle, and a decade of production infrastructure underneath — so the agents I build run somewhere reliable. Three disciplines, one operator.

Three disciplines

One operator, the whole stack.

Each discipline is a working lane, not a job title — the projects on the previous page are where they show up.

Ongoing · core discipline Discipline 01

LLM Infrastructure

The plumbing between a product and every model provider — routing, protocol translation, budgets, and the observability that keeps it honest under load.

  • Multi-provider gateways with routing, failover, and weighted selection
  • Protocol translation across OpenAI, Anthropic, Azure, and Responses APIs
  • Per-key quotas, rate limits, cost tracking, and encrypted credential storage
  • Caching, resilience, and observability with Prometheus and audit trails
  • Self-hosted deployment: single binaries, systemd, Docker, and Cloudflare
  • Model catalogues, pricing data, and virtual aliases for stable routing
Ongoing · core discipline Discipline 02

Agent Tooling

Harnesses, interfaces, and extensions that keep a model useful: code intelligence it can trust, memory that survives a session, and control over the loop.

  • Real-time agent interfaces with streamed tool calls and reasoning visibility
  • Code intelligence: tree-sitter and compiler-backed diagnostics for agents
  • Agent memory: semantic retrieval, consolidation, and cross-session recall
  • Provider and capability plugins: OAuth, vision handoff, and new backends
  • Structured tool design with validation, retries, and error-driven recovery
  • Token and cost observability across sessions, projects, and models
Ongoing · foundation Discipline 03

Platforms & Operations

The layer everything else runs on — containerd orchestration, bare-metal provisioning, and the network operations that decide whether a launch is boring.

  • Game-server platforms on a Rust core with direct containerd integration
  • Multi-node container lifecycle, isolation, and live console streaming
  • Bare-metal provisioning: PXE/netboot pipelines, IPMI, RAID, firmware
  • Proxmox HA clusters, VM/LXC orchestration, and Ceph storage
  • Production service platforms: OAuth, mail ingestion, realtime, and RBAC
  • Network operations, observability, and DDoS mitigation

Toolbox

Languages and platforms I build with.

The short list — what actually ships in the repositories, not what is on a résumé.

Languages 6

  • Go
  • Rust
  • TypeScript
  • Python
  • JavaScript
  • Bash

Data 3

  • SQLite
  • PostgreSQL
  • Redis

Infrastructure 7

  • Docker
  • containerd
  • Linux
  • Proxmox
  • Ansible
  • nginx
  • Cloudflare

Platforms 5

  • Node
  • Bun
  • React
  • Astro
  • WebSocket

Recent work · field note 001

“Built a multi-provider LLM gateway that serves OpenAI, Anthropic, and Responses protocols from one binary — with routing, failover, per-key budgets, and an embedded admin UI. No external database, no runtime dependencies.”
Agent harness ai-gateway · self-hosted, single binary