How I AIHow this PM streamlines 60k-page FDA submissions with Claude, Streamlit, and clever AI workflows
CHAPTERS
- 0:00 – 1:18
Why life sciences innovation still gets slowed by paperwork
Claire and Prerna open with the central tension of modern biotech: world-class science paired with massive regulatory documentation overhead. They frame how cutting paperwork time and cost can directly accelerate access to life-saving treatments.
- •Regulatory submission work is a major bottleneck alongside scientific R&D
- •Paperwork delays can slow vaccine/treatment delivery to patients
- •The episode will focus on using AI to compress cost and timelines
- 1:18 – 1:48
Meet Prerna Kaul: ML-to-GenAI product leader bridging research and approvals
Claire introduces Prerna’s background across Amazon Alexa, Moderna, and Panasonic, setting up why she’s well-positioned to build practical GenAI workflows. The conversation tees up two tracks: FDA submissions and PM communication.
- •Prerna’s experience spans big-tech ML and biotech product work
- •The episode will cover both technical workflow building and stakeholder influence
- •Goal: show end-to-end execution from prototype to usable tool
- 1:48 – 3:45
Sponsor break: CodeRabbit AI code reviews
A brief sponsor segment describes CodeRabbit’s AI-assisted code review workflow and tooling integrations. The pitch emphasizes faster shipping without sacrificing quality.
- •AI-generated PR comments and potential impact analysis
- •One-click fix suggestions and customizable rules (AST grep patterns)
- •Integrations: VS Code, Cursor, Windsurf
- 3:45 – 5:51
The real problem: 60,000-page FDA submissions (BLA) and months of effort
Prerna explains the scale of FDA/regulated submissions, using a Moderna example where a Biological License Application can reach ~60,000 pages. She quantifies the typical resource burden—months of work, ~20 specialists, and millions in cost—and why GenAI was worth exploring.
- •BLA/CTD submissions can be tens of thousands of pages
- •Traditional process: 4–6 months, ~20 specialists, millions in spend
- •Motivation: reduce time/cost and avoid delays in getting vaccines to market
- 5:51 – 7:18
From problem statement to prototype: treating Claude like a software engineer
Prerna shows how she began by giving Claude a concise product-style pitch, including why it matters and what the demo should accomplish. Claude returns not only code but also setup instructions and a demo narrative—accelerating early-stage product definition.
- •Start with requirements and a PM-style “why/what” pitch
- •Claude produced setup docs, task breakdown, and runnable code
- •Claude even generated a demo script/pitch narrative unexpectedly
- 7:18 – 10:12
Production constraints: strict XML structure + safe handling of clinical data
They identify two non-negotiable requirements for real regulatory use: generating structured outputs in a strict XML-based format, and detecting/redacting PHI in clinical data. Prerna explains why Claude’s safety/alignment reputation mattered for this domain.
- •CTD/BLA documents require strict structured formatting (XML-based)
- •PHI detection/redaction is essential for compliance and patient privacy
- •Model choice influenced by safety, alignment, and regulated-industry needs
- 10:12 – 12:03
Model and library choices: medical entity recognition instead of brittle regex
Prerna describes how Claude selected appropriate models for medical named-entity recognition (NER), reducing the need for brittle hand-rolled approaches. Claire highlights how AI removes toil like writing and debugging complex regex patterns.
- •Claude recommended suitable medical NER approach/models quickly
- •Would otherwise require days of research and DS/eng effort
- •Avoids fragile regex-heavy solutions for identifying sensitive entities
- 12:03 – 14:04
Why Streamlit: shipping a usable UI for non-technical stakeholders
Prerna explains that the real unlock is making AI tools accessible to non-technical colleagues via simple UIs. Streamlit becomes the bridge from code to an internal product that teams can actually use, mirroring her other workflow automations (e.g., PRD-to-Jira).
- •Impact increases when non-technical users can use the tool directly
- •Streamlit is a fast way to wrap Python workflows with a web UI
- •Internal tooling examples: PRD-to-Jira conversion and other add-ons
- 14:04 – 15:16
Live demo setup: running the Streamlit app and generating synthetic trial data
Prerna runs the Streamlit app locally and explains the CTD structure and the kinds of clinical-trial inputs it needs. She uses synthetic data to validate functionality while keeping privacy intact.
- •Command-line launch opens a local web app with Claude API integration
- •CTD has multiple modules with specific expectations for study reporting
- •Synthetic data enables testing without exposing real patient information
- 15:16 – 18:53
PHI detection & redaction in unstructured clinical notes
They zoom in on the hard part: PHI isn’t only in neat columns—it’s embedded in free-form clinician notes. The app scans rows to flag and redact dates, names, and other identifiers, showing what a compliant pipeline needs.
- •PHI appears in both structured fields and unstructured notes
- •The app detects and redacts names, dates, and birthdate-like identifiers
- •Traditional PHI pipelines often take months due to high accuracy/compliance bars
- 18:53 – 22:14
Generating the FDA-ready CTD summary and exporting to required formats
After redaction, the workflow summarizes trial data into an FDA-oriented synopsis and generates a structured document output. They highlight that this work is typically owned by medical writers and requires extensive stakeholder iteration—making automation highly valuable.
- •Model produces participant counts, demographics, and study synopsis
- •Output can be downloaded as text/doc plus the required XML format
- •Automation reduces multi-week stakeholder back-and-forth on drafts
- 22:14 – 24:36
Making AI ROI legible: per-operation cost and latency tracing
Prerna adds an operational cost-tracing layer to address stakeholder concerns about scaling token and licensing costs. The UI breaks down duration, token usage, and cost per step, enabling transparent ROI discussions and production monitoring.
- •Stakeholders often push back on AI due to perceived cost risk
- •Cost tracing shows cost per operation, duration, and token breakdown
- •Transparency helps compare alternatives and justify scaling decisions
- 24:36 – 26:42
Measured impact: accelerating vaccine programs and reducing repeated work
Prerna explains how the approach helped real Moderna programs (e.g., cancer vaccine work and combined RSV/COVID efforts) by reducing repeated effort and scaling with the organization. The core value is speed-to-submission and reduced friction across teams.
- •Applied to multiple vaccine programs beyond a single one-off
- •Big wins come from reducing repeated work across many trials/products
- •Time savings can translate into earlier access to life-saving vaccines
- 26:42 – 27:43
Workflow #2: an AI communication coach for high-stakes stakeholder management
The conversation shifts from biotech documentation to a universal PM pain: aligning diverse stakeholders with competing priorities. Prerna describes using Claude as a personalized brainstorming partner for influence and communication strategy.
- •PM effectiveness often hinges on tailoring communication to individuals
- •Claude used as a preparation and brainstorming partner before meetings
- •Goal: faster, more confident planning for ambiguous stakeholder conflicts
- 27:43 – 31:46
Building the coach: Claude Prompt Generator + project knowledge (Carnegie, Austen, persuasion)
Prerna demonstrates creating a reusable, highly structured prompt using Anthropic’s Prompt Generator and iterative prompt optimization. She then “trains” the project with public-domain books (Project Gutenberg) to ground recommendations in persuasion frameworks and literature.
- •Prompt Generator produces structured prompts with defined inputs/outputs
- •Prompt optimization loop improves the prompt based on feedback
- •Project knowledge includes classic persuasion texts and literature for richer guidance
- 31:46 – 33:57
Stress-test scenario: multi-stakeholder conflict two weeks before a major presentation
Claude generates a realistic workplace scenario with privacy and accuracy concerns surfacing right before a key customer moment. The tool enumerates stakeholders and their priorities, framing the core leadership challenge to be solved.
- •Scenario includes compliance, ethics, competitive pressure, and exec expectations
- •Stakeholders are explicitly listed with distinct motivations
- •The tool structures the problem before proposing solutions
- 33:57 – 38:47
Outputs that PMs can act on: strategy options, 1:1 plans, meeting agenda, and back-pocket answers
The coach produces a layered response: situation analysis, recommended strategic posture, suggested leader archetypes, and practical meeting plans. It goes as far as creating minute-by-minute leadership-meeting structure and curveball questions to prepare for.
- •Summarizes and validates the organizational context and core tradeoffs
- •Recommends specific strategic approaches (ethics, compliance, collaboration)
- •Provides stakeholder-by-stakeholder conversation goals and questions
- •Generates an agenda plus “curveball” questions for executive readiness
- 38:47 – 45:13
Lightning round: adopting AI in regulated domains, safety frameworks, and final wrap
Prerna shares advice for peers: identify time drains, delegate via AI projects, and share tools across teams. She outlines a safety-first approach for regulated industries—benchmarking models on ethics, monitoring in production, and having rollback plans—then closes with tips for getting LLMs to ‘listen’ and where to connect.
- •Adoption tip: map energy-draining tasks and delegate them to AI workflows
- •Safety priority: evaluate models on ethics/alignment and known edge cases
- •Operational discipline: offline evaluation + online monitoring + rollback readiness
- •Prompting tip: add emotion/emojis and clear structure when models miss the mark
