Role: Technical Product Manager (Freelance) also operating as de facto Head of QA and a hands-on builder Company: AI/SaaS startup, drug discovery and bioinformatics platform (Seattle) Duration: May 2024 – Present Stack: Python, Claude Code, TypeScript, Cypress, Postman, Docker, Linux

TL;DR

Joined as a freelance Technical Product Manager and Scrum Master. Expanded the role over time to include hands-on building, shipped four production AI agents into the platform's research workflow, built bioinformatics tools manually in Python before leading the transition to an automated tool-builder agent, and rebuilt the team's QA discipline from scratch. Post-launch defects dropped ~30%. The Cypress test suite became fully automated and now runs an evaluation gate against every release.

The situation

The platform supports drug discovery for pharmaceutical research, digital patient simulations, multi-omics analysis, and AI-driven hypothesis and tool generation, served across Free, Academic, Pro, and Enterprise tiers.

When I joined, the engineering team was strong and the science was novel, but the delivery and intelligence layers needed work. There was no formal QA process; engineers tested locally and shipped to production. Bugs were discovered post-deployment, sometimes by paying users on Pro and Enterprise tiers. Separately, the platform's AI surface area was growing and required dedicated product ownership, researchers needed agentic workflows that could match the speed of their experiments, not static tools.

My original scope was sprint coordination and product specs. Within the first quarter, I'd expanded into two parallel tracks: (1) rebuilding QA discipline and (2) directly contributing to AI agent development.

What I did

Track 1: AI agent development

I contributed to four production AI agents now integrated into the platform's research workflow. Each one targets a specific friction point in the researcher's daily work:

• Literature Search Agent: queries multiple scientific databases against a user prompt and returns ranked, relevant articles. Replaces a manual cycle that previously took researchers hours per session.
• Hypothesis Generation Agent: synthesizes user queries with prior experiment context to generate testable hypotheses for researchers. Bridges the gap between "what does the data suggest?" and "what should we test next?"
• Data Analysis Agent: analyzes experimental outputs, returns generated Jupyter notebooks (.ipynb) and visualizations. The output is consumed downstream by the platform's scientific paper writer (which I also contributed to).
• Automated Bioinformatics Tool Builder: agent that generates analysis pipelines on demand. I built the original tools manually in Python before championing and leading the transition to the automated builder, which only worked because I'd already mapped the full surface area of what researchers actually need.

I use Claude Code to ship features end-to-end alongside engineers, not just specify them, the agents above include code I've directly written or refined.

Track 2: QA discipline from scratch

In parallel, I rebuilt the team's testing practice in three layers: