PROJECT 02 / AI automation / Field service
Taskora.
Make operational work visible, repeatable, and easier to improve.
The project
Taskora helps field-service teams automate administrative work and check compliance in the systems they already use. My contributions span the React dashboard, Chrome extension, reporting APIs, PostgreSQL, and agent runtimes.
My contribution
I built the first compliance reporting surface and expanded the product through workflow editing, historical evaluation, browser automation, and runtime reliability. The supplied work record covers 173 merged pull requests between October 2025 and September 2026.
Project details
- Built ClickHouse-backed analytics with account-scoped filters and drill-downs from group and role to user, task, and individual step.
- Added step feedback, manual compliance overrides with reviewer reasons, task-history indicators, and report exports.
- Consolidated task retrieval from more than ten queries to one RPC, batched status polling, and introduced local JWT verification and session caches.
- Built a database-driven model registry, per-step model selection, OpenAI Responses API integration, and priority-based model failover.
Bulk work inside the browser
Built ServiceTrade bulk compliance processing in a Chrome MV3 extension: background tabs capture record data, five workers process the queue, and users see progress and can cancel. Follow-up work covered side-panel reconnects, service-worker lifecycle, record-level compliance indicators, and a typed background dispatcher.
Improve a workflow before publishing it
Built a step editor with draft autosave, comparison, version history, restore, keyboard navigation, and atomic publishing. Streaming AI chat proposes edits that users can apply to a draft and evaluate against captured historical data. Added backtesting, feedback context, and a bounded tool-calling loop.
Agents that fit the product
Worked on Docker-based tenant agents and a second runtime on Cloudflare Workers and Durable Objects. Added persistent Browserbase contexts, browser replay, file round-trips, and runtime routing. The initial Cloudflare delivery was a gated QA release. Also addressed durable scheduling, approval-before-scheduling, tenant timezones, email previews, playbook versioning, and concurrent-run controls.
Correctness behind the dashboard
Corrected ClickHouse CDC queries that counted multiple versions of the same rows and inflated compliance metrics. Moved document extraction to Mistral OCR with R2 staging, tightened account authorization and API security policies, bounded logging schemas, and repaired database migration and function-permission drift.
Delivery and operations
Established QA deployment paths and CI checks across the dashboard, APIs, and extension. Maintained Jest, React Testing Library, and Playwright coverage; documented APIs with OpenAPI; and improved request logging, trace context, signup qualification, and operational review tooling.

