Josh Symonds

• Marvin: Created Marvin, a company-wide agent platform in Slack for code review, incident investigation, and internal workflows: 116 users across engineering, support, and operations, about 50 weekly active.
• Orchestration & Gateway: Designed Go/Kubernetes orchestration for ephemeral agent workers on a Nix-built image, with per-user OAuth credentials, Cedar tool policies, and versioned model routes with one-step rollback. Launched the company’s LLM gateway: SSO-issued keys, request logs, per-person spend, and zero-retention providers only.
• Code Review: Shipped multi-stage AI code review with a verifier pass that checks findings and drops nitpicks: 2,100+ reviews, 87% positive (300+ ratings).
• Evaluation: Built an evaluation harness that replays production reviews against candidate models, with frozen dev/holdout cases and a version-pinned LLM judge. Blind review found the original verifier wrong on 37% of disputed calls. Moved review to open-weight models: cost per review fell 16× in a month, and median review time halved.
• Incidents: Automated investigation of 2,170+ alert threads, gathering evidence from kubectl, BigQuery, and logs before escalating or opening a merge request with the fix, mergeable as written 95% of the time.
• Agent-Ready Monorepo: Made a large monorepo agent-ready with docs, agent skills, and on-demand context, shared by Marvin and engineers’ own Claude Code sessions.

• Shipped an AI gaming companion for 14 games, in OpenAI’s connector directory: sandboxed WASM save parsers on the player’s device, so raw saves never leave it, and MCP tools and the chat loop on Cloudflare Workers behind OAuth 2.1.
• Built its eval system before choosing a model: 295 scenarios × 3 runs from production logs. A 13-model bake-off ranked an open-weight model first, ahead of frontier models costing 14× more. Replaced LLM-as-judge with a decider model (Jev) validated on held-out pairs, which picked the stronger model in 95% of pairs vs 88% for the old judge (p = 1.9e-17).
• Built a Go service pooling warm LuaJIT processes so agents can test builds against Path of Building’s calculation engine (40 ms median).

• Released Gambit, a Claude Code plugin that takes tasks from spec to release with test-evidence gates. Tiltyard replays real repository tasks to score models and agent harnesses.

• Run its AWS/EKS platform and GPU inference for its ML product (Karpenter, KEDA); wrote its SOC 2/GDPR controls.

Original author of Dynamoid, the Ruby ORM for DynamoDB (13M+ downloads). nix-config: NixOS and Home Manager for a nine-host fleet (837 stars).

• Platform: Built the control plane and deploy tool (approval-gated, about 5,800 deploys in 2025), the end-to-end test framework, and the CLIs; ran 9 AWS accounts and 10 EKS clusters.
• Edge: Built the industrial IoT edge OS with zero-touch provisioning and over-the-air updates, and led the migration from k3os to Talos Linux as the fleet grew to 123 production deployments.
• Security & AI: Led SOC 2 Type 1 and presented an IIAR paper on industrial cybersecurity. Set the LLM usage policy, ran weekly “Cooking with Claude Code” workshops, and stood up self-hosted LLM inference for internal search.

• Kubernetes operators and multi-region failover on AKS, custom Terraform providers, and NIST CSF and ISO 27001 programs that passed bank audits.

• Employee #8. Built M1’s infrastructure from launch to nearly $1B in assets: Chef, then self-run Kubernetes, then EKS. Ran security through SOC 2 and SEC, FINRA, and bank audits.