Scale Data

Founder / Principal. 2024 to present.

Stout Tent

Stout Tent's operations application had become difficult for crews to rely on: schedules and travel records were unreliable, and duplicated data made it hard to establish what was current. I interviewed the people using it and investigated the underlying data, using Airbyte to bring MongoDB and Google Sheets into Snowflake for interim reporting. Those reports gave the business something it could use while I developed the replacement with coding agents.

A colleague and I built the proof of concept before I took over implementation for the April 2026 festival deadline. The application covers crew scheduling, travel, equipment logistics, and operational data entry, including the cleanup of duplicated legacy records. I continued interviewing users while they were running events so changes could be checked against actual operations.

The work includes AI in the product and its hosting platform, as well as in development. Product decisions still come from understanding what people need to do. For example, operations needed to deal with missing travel details before an event, so I built a report that ranks events by the missing information and lets staff address or waive each gap in a detail panel. Event-specific subscriptions let them receive the report for the work they are responsible for.

Django, Django Ninja, React, TypeScript, PostgreSQL, Bryntum Scheduler Pro, Pydantic, n8n, Redis, GCP

The platform

In January 2026 I designed and built a shared application-hosting platform on GCP and migrated existing client applications from AWS. Client isolation and role-based access sit alongside separate development and production environments, with backups, monitoring, and logging managed as part of the platform. Infrastructure is defined in code; delivery uses GitOps overlays, images pinned by commit SHA, and validation before apply.

Development with coding agents

The development tooling gives agents work they can finish within a defined scope. A chunker breaks a specification into bounded pieces, each assigned an isolated worktree so parallel changes do not overwrite one another. A swarm coordinator keeps track of the overall intent while worker agents write the feature code, and wt gives each worker its own environment and branch.

Coding agents have development access within client boundaries, with production credentials and promotion outside their control. I added autoscaling, self-hosted Kubernetes runners, and Git-based deployment workflows to handle the volume of changes, together with automated testing and scheduled integration checks. Review of generated code is part of that process, but it cannot establish on its own whether the application still works.

Errors to a draft change

A colleague and I built CCO, a repair workflow that takes frontend and backend errors from GlitchTip and classifies them by severity and application domain. It invokes a coding agent to investigate the source and propose a repair as a draft GitHub pull request for review. I extended it with model options through Vertex AI and a second review stage for recurring fixes and related defects; in-app support requests can enter the same workflow.

ArtsEd West

I cofounded ArtsEd West and have been its Business Manager and Technology Director since June 2025. I used AI to develop membership and payment workflows around Stripe and Google Workspace, along with a publishing workflow the other cofounders could use without writing code. They submit website changes through Google Sheets and review the results in a nonproduction environment before publication. The inaugural ALL AZ Summit took place on 11 September 2026.

Current development work

Not a customer system in production.

The development workflow already creates temporary environments for reviewing changes. I am working toward presenting a proposed repair directly to the person who reported the problem, so they can try it and ask for revisions. That acceptance loop is still to be built.

The ability to generate more code increases the importance of deciding which changes deserve to exist. If engineers become responsible for reviewing a growing queue of plausible implementations while the people asking for them remain unable to try the result, much of the supposed gain may have moved into someone else's workload. I want to understand that whole process before treating generation speed as a useful measure of development.

I also do not think a successful test run is enough to decide that a repair helped its user. The tests can establish behavior we already knew to check; the person trying to complete the work may reveal a misunderstanding in the request or the proposed fix. That person needs a meaningful place in the workflow, with human responsibility for what reaches production and a way to ask for another attempt.