Who I am

My name is Kevin Smith. I’ve worked as an engineer and engineering leader for over 20 years, at places like Chef, CircleCI, AdRoll, and Digital Realty, with a couple of founder stints mixed in. I still write code — you can’t lead engineers if you’ve lost touch with the work.

Right now I’m focusing on AI and its impacts. Lately I’ve been designing agentic workflow systems and trying to figure out what engineering organizations look like when the work itself is changing underneath them. I write speculative fiction on the side, which has turned out to be surprisingly useful for thinking about all of this.

What I’m working on now

  • A goal-conditioned mixture-of-experts neural network, inspired by prefrontal cortex theories. Qwen3 embeddings as differentiable goal signals driving MoE routing and FiLM conditioning, running locally on an RTX 3090. The current puzzle is catastrophic interference in compositional goal training.
  • Agentic workflow architecture — composable typed state machines and uncertainty protocols for how agents should handle ambiguity, partial information, and cascading failures. Elixir/BEAM is the target runtime.
  • Chamacs a terminal-based LLM chat client in Go with a semantic memory system I built from scratch (Memorit). Hybrid vector/conceptual/keyword search constructs relevant context per interaction — the point is to stop leaning on ever-larger context windows.
  • nib, an open-source CLI for fiction writers who think in terminals. Everything is plaintext — Markdown scenes, YAML characters, CSV continuity — so git actually works on your manuscript. AI features (continuity checking, style-aware critique, talking to your own characters) are opt-in and pluggable across backends. I used it to draft a 90k-word novel.
  • Wonda, a multi-agent simulation system for fiction writers. Characters as autonomous LLM-driven actors with traits, goals, and dual memory; MCP servers for perception and action. It generates raw material, not finished prose.

Selected past work

  • Chef (2011–2014). Grew product engineering from 8 to 50 and created a UX practice from scratch. Led the rewrite of Chef Server from Ruby to Erlang — shipped two months early, took per-server capacity from ~100 nodes to 30,000, and unlocked enterprise deals with Facebook and Riot Games.
  • AdRoll (2017–2019). Led the Real-Time Bidding and Personalization teams. Erlang bidders handled ~7 billion requests/month across 12+ ad exchanges. Cut production AWS spend 40% with a spot instance migration.
  • CircleCI (2019–2020). Ran three teams (~30 engineers) through a hypergrowth stretch. Scaled the Mac build fleet from 120 to 300+ nodes. Established a new entry-level EM role and mentored the first cohort into it.
  • Digital Realty (2021–2022). Led 16 engineers across two teams to ship Digital Realty’s first global network automation product on deadline, reporting to the CTO.
  • ShareFile (2023–2025). Got tired of waiting for corporate to figure out an AI strategy and built agentic tooling on top of Block’s Goose with Claude and ChatGPT. Product and engineering management teams adopted it into daily workflows.
  • Operable (2015–2017). Co-founded the company and built Cog, a chatops bot with RBAC, first-class environment management, and Docker integration. Elixir core, Go for Docker, a C/C++ Markdown engine for the templating.
  • Collabchek (2024–2025). My first project under the Poiesic Systems umbrella. I shelved it when agentic coding made the product’s core assumptions obsolete faster than I could adapt them.

What’s next

My ideal engineering leadership role is with teams building seriously with AI and comfortable working in genuinely unsettled territory.

If that sounds like you, let’s talk.

What You’ll Find Here

  • Technical posts on things I find interesting
  • Analysis of problems I can’t stop thinking about
  • Random content as the mood strikes