2025 — Today · Blockchain
Software Engineer
The densest year of my career: a great deal of work, a great deal of experimentation, and nearly all of it AI.
That meant two things. I got to build genuinely interesting AI products. And I got to use AI to work on parts of the stack I would not normally have touched — infrastructure, back end — which turned out to be as significant as the products themselves.
The products were a prompt-to-deployed-app studio, a coding agent that writes, validates, previews and publishes, a shared MCP server other agents consumed, and an agent with durable sessions, typed tool parts, and an AST-validated SQL tool. The surfaces happened to be Farcaster. The work is the agent loop and the Next.js it runs in.
The tooling moved quickly all year. Eve is the one I am betting on — I built the Clanker World agent on it, and today I would not build an agent without it. I would stop short of calling it a ten-year bet. The ground is moving too quickly now for anyone to promise a decade.
Was a primary contributor to Neynar App Studio, an AI product that takes a prompt and returns a deployed app (a Farcaster mini app in this case)
Led the Snap Builder, a coding agent that writes, validates, previews and publishes Farcaster Snaps — built on OpenCode and running on a full Cloudflare stack of Workers, Containers, Durable Objects, R2 and KV
Led the Clanker World agent: a reusable core runtime with matching React and React Native toolkits, typed generative UI, durable Postgres-backed sessions that survive a refresh, and an authenticated product integration
Built the agent's SQL tool — a read-only Postgres engine that parses every query into an AST and validates it before execution, under least-privilege roles with verified TLS and bounded results
Built the corpus management system behind the agent's knowledge — a curated data corpus, isolated Notion knowledge sources, and generated data contracts verified in CI
Built the MCP server other agents consumed — one tool surface a variety of different agents could work against instead of each one growing its own
Designed and shipped about half of the Farcaster Snap SDK, with one other engineer: a versioned declarative UI protocol with typed validators, matched React and React Native renderers, browser and native emulators, documentation, and the release pipeline behind them
Integrated Snaps across the Farcaster clients on web and React Native — composer, feed embeds, tap-to-lift, wallet actions, history and forking, and the analytics behind all of it
Built the backend behind Snap authoring — the service layer the coding agent ran against: asset upload, proxy and analysis exposed as agent tools, build state streamed over SSE, history and draft previews, an Azure-backed safety classifier, feed promotion and URL blocklists
Guided agents through provisioning the infrastructure these products ran on: Terraform for Azure OpenAI, AWS EKS, Postgres, ElastiCache and ECR, alongside Cloudflare R2 storage, Doppler and Vercel secret paths, and Datadog
Made significant improvements to Neynar's developer portal, particularly around user sessions: server-managed JWT authentication that worked properly with server rendering, alongside onboarding, plan selection and direct Stripe card payments beside the existing crypto paths
Built the packages the rest of the frontend shipped on — @neynar/ui and its React foundation, theming, Storybook documentation and npm release workflows — plus a code generator that turns an onchain contract address into type-safe viem clients and wagmi hooks
Prototyped a multi-agent orchestrator that ran agents and workflows, interactively or unattended, and let agents create and delegate work to other agents