How I AIHow this PM streamlines 60k-page FDA submissions with Claude, Streamlit, and clever AI workflows
CHAPTERS
Meet Prerna Kaul: GenAI PM bridging big-tech ML and life sciences
Claire introduces Prerna Kaul and her background across Amazon Alexa, Moderna, and Panasonic. The episode frames two themes: accelerating regulated FDA submissions with GenAI, and using Claude as a communication coach for PM stakeholder management.
Why FDA submissions are a bottleneck: 60,000 pages, months, and millions
Prerna explains the scale and pain of creating a Biological License Application (BLA) and related submission artifacts. The work can require ~20 specialists over 4–6 months, slowing delivery of life-saving vaccines and treatments.
From PM requirements to working prototype: briefing Claude like a software engineer
Prerna starts by giving Claude a problem statement, pitch narrative, and demo goals—treating the model like a full-stack engineer. Claude returns not only code but also setup instructions and even a demo script, accelerating early iteration.
Two non-negotiables in regulated workflows: strict structure + PHI safety
Prerna identifies the core requirements for production readiness: generating strictly structured, XML-based outputs and reliably detecting/redacting PHI. These constraints drive model/vendor selection and the overall system design.
Modeling choices for medical entity recognition and redaction—without months of R&D
The conversation highlights how Claude helped select appropriate approaches for medical named-entity recognition and PHI redaction. Prerna notes that historically this would take multiple specialists and months to get right across varying trial datasets.
Making it usable: why Streamlit is the ‘unlock’ for non-technical stakeholders
Prerna explains that building a simple UI is what turns a personal automation into an organizational tool. Streamlit enables quick deployment of Python workflows so medical writers, compliance partners, and other non-technical users can operate the system.
Live app walkthrough: generating synthetic trial data and previewing clinical notes
Prerna runs the Streamlit app from the command line and demonstrates the workflow using synthetic data. The demo mirrors real trial structures: tabular participant data plus free-form clinician notes where PHI may be embedded.
Detecting and redacting PHI in unstructured clinical notes
The demo shows the system scanning records to identify names, dates, and other PHI within narrative text. The redacted output preserves medically relevant content while removing identifying details suitable for downstream reporting.
Generating the Common Technical Document (CTD): summaries, stats, and XML output
With data de-identified, the workflow produces CTD-style modules that include participant counts, age ranges, methodology summaries, and medically relevant terminology. Prerna contrasts this with medical writing workflows that can involve weeks of stakeholder iteration per module.
Proving ROI: tracing operations and showing per-step token and dollar costs
Prerna emphasizes that cost transparency can unblock adoption in organizations skeptical of AI spend. The app shows cost per operation, duration, and token breakdown—useful for budgeting, scaling forecasts, and production monitoring.
Real-world impact: faster submissions, scalable processes, and vaccine timeline benefits
Prerna shares that applying these workflows to major vaccine programs produced meaningful time and cost savings. Beyond immediate savings, the bigger win is creating a scalable system that reduces repeated work as drug portfolios expand.
Switching gears: building an AI ‘Influence & Communication’ coach for PMs
The episode transitions to a second workflow: using Claude to improve stakeholder communication. Prerna frames influence as personal and context-dependent, making it a strong fit for an AI brainstorming partner that can structure preparation.
Prompt Generator + optimization: structured XML prompts and project-based coaching
Prerna demonstrates Claude’s Prompt Generator to create a reusable, structured prompt with role definition, inputs, and output format. She then uses Claude’s prompt-optimization loop to iteratively improve it and installs it as project instructions.
Training Claude on classics: Gutenberg books, Dale Carnegie, and strategy inspiration
Prerna loads public-domain texts (e.g., Dale Carnegie, classic literature) as a project knowledge base to enrich the coach’s suggestions. Claude selectively retrieves relevant excerpts rather than dumping all sources, grounding recommendations in familiar frameworks.
Stress-test scenario: multi-stakeholder conflict two weeks before a major pitch
Claude generates a realistic workplace crisis: privacy and accuracy issues discovered shortly before an important prospect meeting, with stakeholders pulling in different directions. The coach then produces a structured plan: analysis, stakeholder 1:1s, and a leadership meeting agenda with contingency questions.
Lightning round: adopting AI in regulated industries, safety priorities, and ‘LLM listening’ tips
Prerna advises peers to identify energy-draining tasks and delegate them to AI workflows that can be shared across teams. She underscores ethics, privacy, and continuous evaluation/monitoring as priority zero in regulated contexts—and shares a practical trick for getting better responses: add emotion and even emojis.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome