Systems that run without me standing next to them.
Three builds — one you can use right now, one running unattended in production, one described but not shown, because employer code stays the employer's. Every claim below is either demonstrable or plainly labeled.
Autonomous AI trading pipeline
An example of how I build — a personal project shown for its engineering. Not a product, and not a trading, investment, or financial service I offer.
Could an AI agent run a disciplined, rules-based research-and-trading process entirely unattended — without a human babysitting it, and without any ability to hide what it did? Most "AI agent" demos run once, supervised, on a good day. This one had to wake itself up on schedule, do real multi-step work, and leave a paper trail a skeptic could audit.
Six scheduled routines (pre-market, midday, close, weekly review, discovery, verdicts) each run as a one-shot agent with a written spec, hard rules it cannot override, and a requirement to commit its full reasoning to a git journal before it finishes. Strategy changes are human-gated — the agent proposes, a person approves, git records.
Every routine can crash and re-run without double-trading or corrupting state.
Decisions, reasoning, and outcomes are committed — the system cannot misremember what it did.
The agent cannot change its own rules. Proposals require sign-off; the change history is versioned.
Two independent flags must both flip before real money is possible. Neither has flipped.
Every run leaves a log and pushes a notification — silence is itself an alert.
Cron drift, API outages, and mid-run crashes have all happened — and been absorbed by design, not luck.
Running unattended daily for over a month: every decision timestamped, versioned, and auditable. Deliberately paper-only while the strategy accumulates a statistically honest sample — the engineering discipline is the demonstration here, not the returns. That's the same discipline I bring to client automation.
Selenium → Playwright migration, enterprise scale
An aging Selenium suite: slow regression cycles, brittle waits, flaky runs that trained the team to distrust red builds — the point where automation quietly stops paying for itself.
Designed and built a Playwright framework from the ground up — resilient locators, auto-waiting, database-driven validation (Playwright + SQL Server) that checks the data behind the UI, wired into Azure DevOps pipelines so release-readiness checks run on every build. Authored 5,000+ automated tests, and I still own and run it today.
Regression cycle time down 35%+. Manual testing effort down ~40%. Suites exceeding 1,200 tests running in CI. When a test fails now, the team trusts it — and I isolate the defect to the exact commit.
This site — an ROI argument you can poke at
Conversations about automation ROI run on vibes. Skeptics have been burned by vendor math that hides the upfront cost; believers can't show the payback curve. Both deserve numbers they can argue with.
An interactive single page where your team's real inputs drive every number — a simulated suite run, a manual-vs-automated race, and a break-even chart that models the scripting cost and a 17% maintenance tax against the savings. Vanilla JavaScript, zero dependencies, no tracking, loads in under a second.
The site about test automation is covered by its own Playwright suite — functional, accessibility (WCAG AA), and cross-browser checks, with the ROI math itself pinned to hand-computed expected values, running in CI on every push. View source; that's the point.
That's the kind of problem I like.
Test automation, CI/CD, AI systems that have to prove themselves. If you're wrestling with one, I'd like to hear about it.
Let's talk automation