- Ran the load testing after a December 2025 outage. There was no clean data for the peak hour, so I rebuilt it from 15 to 20 days of backlogged submissions and timing patterns from 2024, then simulated up to 3,000 users an hour with k6.
- Built DevKit, a local copy of production the 12-person team uses every day: Docker and Tilt running Keycloak, search, SQL Server, the Go and Java APIs, and Grafana.
- Tracked down SQL deadlocks and Kubernetes resource limits, and walked teammates through avoiding the deadlocks.
Bryce Davidson
Software engineer: backend and applied AI
Victoria, BC, moving to Vancouver
bryce_davidson@icloud.com
GitHub ·
LinkedIn ·
S4MPLE
I build backend systems, and the checks that let a team rely on what its coding agents produce.
I'm a backend developer at EducationPlannerBC, the service students use to apply to BC colleges and universities. Outside work I'm building S4MPLE, a learning app for web, Mac, iPad and iPhone, much of it written by Claude Code sessions running in parallel. I'm moving to Vancouver and looking for engineering work on teams that build with AI tools.
Work
- One TypeScript codebase for web, Mac, iPhone and iPad, with Stripe and Apple billing. In beta with early users.
- An evaluation harness for the agents' instructions. It replays 21 cases drawn from real agent mistakes as headless Claude Code sessions, grades each transcript with rule checks and an LLM judge, and can score two versions of the instructions side by side.
- Tested hooks for the rules that can't be left to judgment, such as refusing a merge until CI passes on that exact commit.
- An MCP server that lets AI assistants read and write a user's notes, with its own OAuth sign-in, scoped access and rate limits.
- Worked through the repair of a corrupted MongoDB database of genetic data, recovering 83% of the affected records.
- Moved external Kubernetes clusters to GitOps with FluxCD, so each cluster's setup is version-controlled.
- Built RadialSearch, a Julia library for variable selection in linear models that spreads the search across many processors.
- Applied deep learning to estimate cell-type makeup in lung-disease samples, alongside master's and PhD researchers.
Writing
How I test the instructions my coding agents follow
In progress. What an eval harness for agent instructions looks like, what it caught on the first run, and where it stops.
Smaller projects
- Fox3 International flight debrief. Built with a pilot-training school's owner: a prototype that replays training flights in 3D from Garmin flight logs, several aircraft at once.
- EventZen AI phone assistant. Answers event attendees' questions, takes RSVPs and texts them a summary.
Outside work
Road cycling, ten to twelve hours most weeks, and backcountry hiking. Before software I worked as a barista and as a photographer in Banff and Victoria.