Traverse Operations

I build operations that run while you sleep.

Traverse Operations is a one-person applied AI and operations practice in Seattle. I design, deploy, and operate agentic systems that do real work overnight and hand a person only the decisions that need one. Every pattern on this page has run unattended, in production, on systems I built and operate.

Start a conversation

What I build

  • Agentic systems and workflow automation

    Autonomous agents, decision-support tooling, and multi-step workflows built with Claude Code, MCP, n8n, and AI-assisted Python and FastAPI.

  • Self-auditing decision systems

    Rubric-based scoring engines that keep a permanent ledger of every decision, re-run a pinned fixture set on a schedule to catch their own drift, and regression-test rule changes before adopting them.

  • Human-in-the-loop operations

    Nightly agents that plan, research, build, and test on their own, and stage anything external, public, or money-spending as a decision card for a person to approve.

  • Operational infrastructure

    Containerized, self-hosted, Tailscale-only. Ephemeral non-root agent containers behind an egress allowlist, scheduled jobs, watchdogs that are silent on success and loud on failure, and backups that are tested.

What Traverse can take on

Every pattern below has run unattended in production on systems I built and operate. The same patterns apply whether the workflow belongs to one person or to an operation with hundreds of people in it. I have run both.

  • Unattended agents

    Agents that work a queue while nobody watches

    Scheduled agents that plan, research, build, and test inside a fixed window, work a task queue until the budget is spent, and stand down on their own when a quota, a deadline, or a failed check says so. Higher-priority work always keeps its reserve.

  • Decision systems

    Scoring engines that can prove they still work

    Rubric-based evaluation of high-volume inputs with a permanent ledger of every decision. A pinned fixture set is re-run on a schedule to catch drift, rule changes are regression-tested before they are adopted, and the audit itself produces the fixes.

  • Human gates

    Automation that knows what it may not decide

    A decision inbox where anything external, public, or money-spending waits as a card for a person to approve. Cards reach a phone with quiet hours. Approvals are logged, denials are logged, and the agent proceeds only on what was approved.

  • Validation

    Finding out whether something should exist

    Structured kill tests and substitute studies with gates written down before the work starts, so a product or a process gets a verdict with a reason rather than a slow fade. Stopping on evidence is treated as a result, not a failure.

  • Local first

    Products that hold as little data as possible

    Applications where the index lives on the user's device and no backend exists to breach. When a product does have to hold data, the design states exactly what is kept, where, and for how long, and the security headers and deletion path exist before launch.

  • Operations

    Infrastructure that reports only when it matters

    Containerized, self-hosted or cloud, with ephemeral non-root agent containers behind an egress allowlist, scheduled jobs, watchdogs that check liveness, staleness, and credential age, and backups that are tested rather than assumed.

How I work

  • A person holds the gates.

    Agents propose, research, build, and test. Anything external, public, or money-spending waits for an approved card.

  • Systems audit themselves.

    Every automated decision is logged. Scheduled audits re-run fixtures against current rules and treat drift as a defect, not a surprise.

  • Stop on evidence.

    A kill verdict with a written reason is a good outcome. Work that should stop gets stopped, and the reason gets written down.

  • Local first, minimal custody.

    Data stays on the user's device or my own hardware wherever the product allows it. When it cannot, the design says exactly what is held and for how long.

  • Quiet on success, loud on failure.

    Watchdogs and alerts fire only when something needs a person. Nothing sends a "still fine" message.

Engagements

Project-based and fractional, remote. The typical shape is an operational process or decision workflow you want automated, audited, and made reliable. I assess it, build it end to end, and hand it off with the monitoring to run it. Most engagements start with a two-week scoped build against one workflow.

Background

More than a decade leading operations at a global technology company, most recently directing a multi-country organization of well over a hundred people that delivered data and human-in-the-loop services for machine learning teams. That work was capacity, quality, cost, and governance at scale. Since 2025 I have applied the same judgment hands-on, as the person who architects, ships, and operates the system rather than the person who specifies it.

Contact

This is a one-person practice. Describe the workflow you want to stop watching, and I will reply with whether it is a fit and what a first scoped build would look like.

Start a conversation