Work with Levelbrook
Senior backend engineering for the part of the system that only misbehaves under load. Ruby on Rails, Go, Python, and Postgres, mostly on applications that already have real users and real data.
Available for contract engagements (C2C or 1099) and open to full-time roles.
Start a conversationOr email levelbrookteam@gmail.com. Tell me what you are building and what is going wrong. You will get a real answer, not a deck.
What I get called in for
The app got slow
Query plans that quietly stopped being true as the table grew, N+1s that only appear for your largest tenant, and dashboards doing arithmetic no index can save. Usually the fix is smaller than expected once it is measured.
The upgrade stalled
A Rails upgrade that has been half-finished for two quarters, a test suite nobody trusts, or a migration everyone is afraid to run in production. Zero-downtime migration work is a large part of this.
Multi-tenancy is leaking
Tenancy holds in the query layer and breaks in background jobs, exports, cached fragments and support tooling. That is where the embarrassing bugs live.
LLM output has to be trusted
Grounding gates, abstain paths, and eval harnesses. The interesting engineering is deciding when the system should refuse to answer, and proving it does.
Concurrency you cannot reproduce
Goroutine leaks, lock-order deadlocks, duplicate-delivery bugs. I have written static analysis and deterministic interleaving tooling for exactly this class of problem in Go.
Deploys are a person, not a process
Docker, Kamal, Cloudflare, small boxes. Making a deploy boring enough that it can run unattended at 3am is usually cheaper than the platform migration someone is proposing.
The evidence
- Fifty-eight engineering essays. Nearly every one starts with a real production problem and a number attached to it.
- tether, a static goroutine-leak
analyzer for Go, benchmarked on the entire Go 1.25 standard library: 21 diagnostics, zero
in non-test code, and a real WaitGroup hang found in
database/sql's own tests. - weave, a deterministic interleaving explorer for Go with vector-clock race detection and failing-schedule minimisation.
- A live Hotwire application you can click around in: a kanban board that syncs across browsers, per-field inline editing, and LLM token streaming over SSE. No React anywhere.
- A loan-servicing triage tool whose case state is derived from an append-only event log, shipped with its own forty-case labelled eval, including the one costly miss.
How engagements usually work
Most start with a paid scoping conversation: I read the code, reproduce the problem, and write down what I found and what I would do about it. That document is useful whether or not the work continues. From there it is typically a fixed-scope project or an ongoing weekly commitment. Contract-to-contract and 1099 both fine; MSA, SOW, NDA and COI are all routine.
Start a conversation