Selected work

Six projects, and what actually happened to them

I mostly work on systems that already exist and are quietly struggling. Below are the engagements I can talk about publicly, with the numbers that came out the other side. Client names are replaced where contracts require it.

Northwind reporting dashboard
2024·Lead engineer·Logistics, 180 staff

Northwind — reporting that runs on one box

A freight operator had grown to 40 dashboards, all of them slow, all of them maintained by one increasingly tired analyst. We deleted two thirds of them and rebuilt the rest against a single Postgres instance with proper materialised views.

  • p95 dashboard load 4.1s → 380ms
  • Infrastructure spend down 62%
  • Nightly rollups: 41 minutes → 38 seconds
TypeScriptPostgresLitestreamGrafana
Atlas route planning interface
2023·Backend lead·Field services, 14 countries

Atlas — route planning at street level

Scheduling engine for technicians who drive between jobs all day. The original implementation asked a third-party API for every leg; we moved the graph in-house and precomputed the expensive half of it.

  • Route queries 1.9s → 84ms at p95
  • Third-party API calls down 94%
  • 40M road segments queried in one index
GoPostGISRedisKubernetes
Relay event pipeline interface
2022·Architect·Marketplace, 2M users

Relay — 40 million events a day, quietly

Replaced a chain of cron jobs and half-written Kafka consumers with one pipeline that has an idempotency model the team can actually explain during an incident.

  • Two pages in three years, both DNS
  • Duplicate events: 0.7% → 0.0002%
  • Replay of any 24h window in under 6 minutes
RustNATSClickHouseOpenTelemetry
Lumen design system components
2023·Design systems lead·Fintech, 4 product teams

Lumen — one token pipeline, three brands

Four teams, four button components, three brand skins. We built the token layer first and let the components fall out of it, which is the only reason the system is still in use two years later.

  • 64 documented components shipped in six weeks
  • Brand theming: 3 weeks of work → one config file
  • Contrast failures in audit: 212 → 0
TypeScriptStyle DictionaryStorybookFigma API
Ferrite index advisor output
2022 — present·Maintainer·Open source, MIT

Ferrite — a Postgres index advisor

A single-binary CLI that reads pg_stat_statements and tells you which indexes are missing, which are dead weight, and what each one costs you in write throughput. Started as an internal script.

  • 2,400 GitHub stars, 61 contributors
  • Running in production at 300+ companies (that I know of)
  • Average 7 unused indexes found per audit
RustPostgreSQLSQL
Quill editorial CMS
2021·Full-stack·Newsroom, 40 writers

Quill — a CMS a newsroom would keep using

The previous editor lost drafts on bad hotel Wi-Fi, which is the whole reason this project existed. Offline-first editing, deterministic conflict resolution, and an image pipeline that respects deadlines.

  • Lost drafts: 3–5 per week → zero reported in 18 months
  • Publish-to-live: 90s → 4s
  • Image pipeline cost down 71%
TypeScriptPostgresS3Service Workers

Outcomes at a glance

The numbers I would defend in a review

Measured outcomes per project
Project Metric Before After
NorthwindDashboard p954.1s380ms
AtlasRoute query p951.9s84ms
RelayDuplicate events0.7%0.0002%
LumenContrast failures2120
QuillPublish-to-live90s4s

Figures are from client-side telemetry and my own load tests. Where a number came from a vendor dashboard I say so on request — I would rather show the query than the graph.

Want the long version of any of these?

I am happy to walk through architecture decisions, trade-offs and the parts that did not work. No slide deck required.