Sudarsh M. Sathianathan

Work

A product I built, and the pipeline I build it with. My full work history is on my resume.

Jul 2025 – present

Simulr

Founder & Product Architect

Context

QA tools check whether an app works. I wanted to test whether people understand it, so I built an AI testing product around configurable personas.

What I did

  • Built an agentic LLM testing loop: accessibility-tree capture, Claude reasoning, Playwright execution and action logging. The persona's decisions are explained in plain English.
  • Designed configurable personas using a JSON behaviour schema and a variance model.
  • Wrote evaluation rules based on usability heuristics, an error taxonomy and WCAG 2.1.
  • Created the UX Confusion Score.
  • Ran product discovery with PMs and founders.

Results

PMs and founders liked the approach. I also learned that teams with existing UX backlogs couldn't take on new findings.

Stack

TypeScript, Python, Claude API, Playwright, Next.js, FastAPI, Supabase

See the Simulr page →

Used to build Simulr

Agent-driven delivery pipeline

Context

Ask one agent to build a feature and write its tests, and the tests can only check that agent's own understanding of the requirement. I built this pipeline so agents prove their work against scenarios they didn't write the code from.

What I did

  • A Claude Code scenario agent turns acceptance criteria into structured scenarios.
  • Scenarios are reviewed and frozen.
  • Separate dev and test agents work on separate branches.
  • The test agent writes Playwright tests from the frozen scenarios, not from the dev code.
  • GitHub Actions runs unit, API and UI checks, plus scenario-to-test traceability.
  • Failures are classified and routed.
  • Pull requests get an AI review.

Results

Tests come from agreed scenarios instead of the implementation, and traceability shows which scenarios have a test. I wrote up the full design, including what still needs people.

Stack

Claude Code, GitHub Actions, Playwright

Read the article: Agent-Driven Continuous Deployment: How AI Agents Take a Ticket to Production Without Breaking It →