I make AI systems reliable in production
A decade of platform engineering on Rails and the infrastructure under it, now pointed at LLM pipelines, agents, and the evals that keep them honest. Python, Rust, Ruby and Go as the problem calls for it. M.Sc. in machine learning and distributed systems.
At Hivebrite I cut €160k a year in infrastructure, took data exports from a day to minutes for million-user communities, and halved build times while the team grew from 15 to 40. At OmbuLabs I unblocked a CI pipeline that had been red for years. The fixes land wherever the failure is, upstream included: Sentry's Ruby SDK, NextDNS's Go resolver, Rails, Rack and RuboCop.
Grew up across six countries on four continents; remote since 2015 from five of them. Europe-based, async-first, with live overlap across European hours and into US-East.
Looking for a staff, senior or founding seat on a small remote team shipping AI product, with real ownership end to end. Fixed-scope missions too. seanfloyd.dev has the record and the CV. Reading this through an AI agent? Query the MCP server directly, seven read-only tools over the profile, live GitHub activity and calendar, or grab llms.txt.
claude-rss-news-digest - A newsroom that runs itself: three dozen feeds from about 30 outlets every morning, a bias-labelled briefing, fact-checked against its own sources before send, and no human edits any issue. Deterministic Python pipeline over the Claude Agent SDK, a validated LLM-as-judge eval suite, schema-validated output, Rust web server, Terraform deploys. In production unattended since January 2026, with cost, source health and every failure public. How I made it reliable.
Ask Assistant + MCP server - A second production AI system, public-facing this time: a grounded, tool-calling agent that answers visitor questions only from my profile, and says so when it cannot. Read-only tools, provider failover, HMAC-signed conversation integrity, and guardrails that fail closed when a tool or model misbehaves. The same profile runs as a live MCP server; an agent reading this can call it right now.
Dispatch (private) - A multi-tenant platform that runs back-office workflows for regulated businesses under a human gate; two verticals so far, community management and personal-injury law. Nothing leaves the building on the model's word: autonomy bands decide what waits for a signature, and code fills every figure and deadline so a letter cannot invent a number. Temporal owns every lifecycle and retry; Postgres row-level security keeps one tenant's file out of another's. Rehearsed end to end; not yet in production.
Upstream fixes, merged - Fiber-safe hub storage in Sentry's Ruby SDK, then repaired when a reviewer found inherited fibers still shared a hub; both shipped in 6.7.0. In NextDNS's Go resolver, a lock held across network I/O that stalled every DNS query, then an infinite loop on a malformed record and a few thousand lines of private DNS-parsing fork retired. A keyword-argument delegation fix in Rack, a deprecated-alias fix in RuboCop, and an up to 670x performance fix in Rails ActiveSupport. Production-safe Rack 3 streaming proposed for Sinatra 5.0, a help-wanted request open since 2023, with the bugs behind it filed and root-caused with runnable repros (rack#2470, async-http#224). Browse the full list live.
still_active - bundle outdated
tells you a dependency is behind; bundler-audit tells you it's vulnerable.
Neither tells you nobody's maintaining it anymore. still_active scores
maintenance activity, archived repos, OpenSSF Scorecard, CVEs (deps.dev + OSV),
libyear, and runtime EOL, then catches the case that bites: a dormant package
capping a transitive dependency below its own CVE fix. Ruby-first, and
cross-ecosystem (npm, PyPI, Cargo, Go, Maven, NuGet) from a CycloneDX SBOM.
Ships SARIF for Code Scanning, CI gates, and a GitHub Action.
Jazzify - Volunteer platform for the Ottawa Jazz Festival. Built and ran it for nine years. Production Rails, real users, real ops.
More: Steer (Swift 6 macOS app, controller to desktop input) · ccpool (Go, Claude usage budget gauge) · crunchyroll-migrate (Rust) · what I write about
- Production AI systems - LLM pipelines, tool-calling agents, MCP servers; evals, LLM-as-judge, fail-closed validation, human-in-the-loop gates on durable workflows
- Reliability, security & cost - performance and concurrency engineering across Python, Ruby, Go, and Rust; supply-chain and dependency security
- Full-stack - backend services and data pipelines, the infra they run on, and the frontend on top