How I use it

I work across the complete agent loop: Playwright and Chrome CDP for control, DOM, screenshot and accessibility-tree context for perception, set-of-marks scanning for grounding, and ReAct-style planning for the next action.

Changes are evaluated on fixed episodes with paired seeds, exact McNemar tests, per-arm cost tracking and a pre-committed rule to revert anything without a positive pooled result.

Evidence in the work

• Moved a fixed 45-episode MiniWoB++ regression set from 33/45 to 43/45 through seven paired experiments.
• Kept three general-mechanism wins and cleanly reverted four candidates; every retained change received a regression test.
• Established a fixed-slice WebVoyager baseline and rejected all three architecture candidates when they missed the non-regression gate.

How I use it

I build vertically: interface components, server routes, authentication, relational data, integrations and deployment are treated as one product path rather than disconnected specialties.

Recent work includes Next.js and React interfaces, Node.js REST APIs, Python agent services, MySQL and PostgreSQL data, Clerk authentication, Twilio messaging and webhook-based integrations.

Representative delivery

• Designed a frozen versioned API so web and iOS surfaces could consume one backend through fail-closed adapters.
• Rebuilt a responsive dashboard design system across two themes and two breakpoints while its 291-test suite remained green.
• Built Bridge2AI’s import, schema, prompt and dual-output pipeline end to end for an external industry client.

How I use it

I trace failures across boundaries rather than stopping at the visible symptom: browser state, extension behavior, WebSocket transport, service routing, operating system and deployment environment.

For browser automation, that means controlling the user’s logged-in Chrome through a Manifest V3 extension and CDP while preserving focus, isolating per-job state and failing honestly when the surface is unavailable.

Representative delivery

• Diagnosed transport races, WebSocket fragmentation, hardcoded surface routing and focus-stealing extension behavior.
• Restored background automation with up to three concurrent jobs and a dedicated tab per job.
• Configured and worked across Railway, Vercel, Supabase, Clerk, Twilio and Expo/EAS production services.

How I use it

I work with data as a product constraint: explicit schemas, traceable imports, scoped queries, truthful missing values, retrieval budgets and operational backups.

The range spans application databases and pipelines through self-hosted storage, private networking and routine administration.

Representative delivery

• Imported spreadsheet and CSV data into MySQL, then translated natural-language questions into SQL for Bridge2AI.
• Worked with PostgreSQL through Supabase, SQLite plus embedding retrieval, and a Neo4j-backed memory visualization.
• Deployed and maintain TrueNAS with ZFS storage and snapshots, file sharing and Tailscale remote access.

How I use it

Ariadne Leather joined physical craft with the full operating path around it: workshop tooling, product photography, catalog management, hosting, orders, shipping and customer communication.

I use OpenSCAD for practical workshop templates and jigs, and I treat the storefront as an operated system rather than a decorative web project.

Representative delivery

• Built and administered a WordPress and WooCommerce storefront with MySQL, payments and self-managed hosting.
• Integrated Canada Post live rates, SMTP delivery, automated receipts and review workflows.
• Handled product photography, fulfilment, support and the physical leatherwork behind the catalog.