How Prerna Kaul Automates 60,000-Page FDA Submissions and Coaches PMs with AI

In this ep, I sit down with Prerna Kaul. She’s a product and platform leader who has spent years turning cutting-edge machine learning research into products you use every day at companies like Amazon Alexa, Moderna, and now Panasonic Well.

Prerna works in the high-stakes world of life sciences, where new vaccines and treatments are paired with mountains of regulatory paperwork. And I do mean mountains—we're talking about documents that can run up to 60,000 pages, take a team of 20 specialists up to six months to complete, and cost millions of dollars. The pressure is huge, because any delay in this process means a delay in getting life-saving treatments to the people who desperately need them.

In this episode, Prerna shares two really different but equally clever workflows. First, she walks us through exactly how she used Claude to build a production-ready application that automates the process of creating these massive FDA submission documents. She shows us the whole process, from a simple prompt to a working tool that has a real-world impact. Then, she shows us something just as useful: how she built an AI-powered communication coach, trained on the works of Jane Austen and Dale Carnegie, to help product managers navigate the complex world of stakeholder management.

Whether you're a developer in a highly regulated industry or a PM trying to align your team, Prerna's practical approach is a great blueprint for using AI to solve your own complex challenges—both technical and human.

Workflow 1: From Prompt to Production - Automating FDA Regulatory Documents

The first workflow Prerna shared tackles a huge problem in the pharmaceutical industry: compiling the Biologics License Application (BLA), a document that can reach 60,000 pages. This process is a massive bottleneck, requiring so much manual effort, time, and money, all of which delays getting critical treatments like Moderna's cancer and RSV vaccines to market. Prerna saw how generative AI could completely change the way this work gets done.

Step 1: Prototyping with Claude as Your Co-founder

Like any great product manager, Prerna started with the requirements. But instead of writing a traditional PRD, she treated Claude as a software engineer and collaborator. She gave it a high-level prompt that outlined the problem, the business impact, and the desired outcome.

"The thinking I had in mind is that Claude is a software engineer, and I'm talking to them and trying to tell them why it matters like any good PM would, and trying to tell them what end product we wanna produce as a result."

The results were incredible. Claude went way beyond just understanding the request—it generated a complete project plan. This included:

• A detailed, step-by-step set of instructions in a Markdown file.
• A list of the tool's core capabilities.
• A full narrative script for how to run a demo of the product.
• The complete Python code to get started.

This single step, which would have normally taken a PM and engineering team at least a week of planning, was completed in minutes.

Step 2: Building the Core Logic for PHI Redaction

A critical requirement for this tool was handling sensitive clinical data. It needed to be able to structure the data correctly and also detect and redact any Protected Health Information (PHI). This is a complex task, especially with unstructured data like clinical notes, and often involves specialized data scientists and months of model development.

Prerna was amazed that Claude, with very little domain-specific instruction, immediately identified the correct models for medical named entity recognition and generated the necessary code. It even handled the tedious task of writing complex RegEx for pattern matching—a task I think many of us have lost days of our lives to!

Step 3: Democratizing the Tool with a Streamlit UI

An AI tool is only really valuable if the people who need it can actually use it, and many of them are non-technical. So, Prerna’s next step was to wrap her Python script in a simple user interface using Streamlit, an open-source framework that makes it incredibly easy to build and share web apps for machine learning projects.

She ran the application directly from her terminal, and in seconds, a local web app was up and running, ready for her colleagues to use.

# Example command to run a Streamlit application
streamlit run app.py

Within the app, she could first generate synthetic clinical trial data. This included both structured, tabular data (patient ID, ethnicity, etc.) and unstructured, free-form text in the 'Clinical Notes' section, where sensitive PHI could be hiding. After generating the data, she simply clicked a button to "Detect and Redact PHI," and the application scanned everything, flagging and removing dates, names, and other personal identifiers.

Step 4: Generating the Document and the "Magic" XML Button

Once the data was clean, the final step was to generate the Common Technical Document (CTD), a specific module within the BLA. The app summarized the clinical trial data, participant statistics, and methodologies into a format that the FDA requires.

The app provided two download options: a standard text file and, most importantly, the proprietary XML format required for submission. Seeing that "Download as XML" button was a huge moment.

"I cannot tell you how excited I was to see this button."

Step 5: Gaining Buy-in with Transparent Cost Analysis

We all know one of the biggest hurdles for adopting new AI tools internally is the perceived cost. To get ahead of this, Prerna built a cost analysis feature directly into the app. It tracks the token cost and duration for each operation (e.g., PHI redaction, document generation), providing clear, transparent data on the tool's ROI.

This allowed her to say, for example, "PHI redaction costs us X cents per patient," directly comparing it to the much higher cost of manual review or building a bespoke ML model from scratch. This transparency was crucial for getting stakeholder buy-in.

Workflow 2: Your AI Communication Coach for Stakeholder Management

For her second workflow, Prerna moved from a big technical problem to a tricky human one: stakeholder management. As product managers, we’re always working with different stakeholders who have their own priorities, concerns, and communication styles. Getting everyone aligned is one of the most important—and hardest—parts of the job.

Prerna built an AI-powered coach to act as a brainstorming partner, helping her prepare for high-stakes meetings and develop effective communication strategies.

Step 1: Crafting the Perfect Prompt with Claude's Prompt Generator

This workflow began in the Anthropic Console, where Prerna used a feature called the Prompt Generator. She gave it some basic instructions, and it created a sophisticated, highly structured master prompt for an "Influence and Communication Coach." The prompt uses an XML-like format to break down the problem, define the inputs (situation, knowledge base), and specify the desired output format (analysis, strategy, play-by-play).

Step 2: Training Your Coach on the Classics

With the master prompt created, Prerna set up a new project in Claude and pasted the generated prompt into the project's instructions. The really clever part is what she did next: she uploaded a knowledge base of classic literature and non-fiction focused on persuasion and human interaction. Sourced from Project Gutenberg, her training data included:

• How to Win Friends and Influence People by Dale Carnegie
• Works by Jane Austen
• Other books on tactics and persuasion

By grounding the model in this timeless wisdom, she gave her AI coach a deep understanding of human dynamics.

Step 3: Simulating a High-Stakes Scenario

To test the coach, Prerna asked it to generate a realistic, challenging scenario. Using Claude Sonnet 4, the model created a situation where a PM at a healthcare AI startup discovers major data privacy and accuracy issues just two weeks before a critical presentation to a major prospect. The scenario included a full cast of stakeholders with competing concerns: the CEO, the Head of Sales, the Head of Clinical Data, and the Chief Legal Officer.

Step 4: Generating a Strategic Playbook

With the scenario set, Prerna simply asked the coach to proceed. Drawing on its instructions and knowledge base, the AI generated a comprehensive strategic plan. The output was incredibly detailed and actionable:

• Situation Analysis: A breakdown of the core challenge, validating its understanding of the nuances.
• Inspiration from Tech Leaders: It offered perspectives on how leaders like Satya Nadella or Andy Jassy might approach the problem, grounding the advice in real-world examples.
• Communication Principles: It extracted relevant principles from Dale Carnegie's work.
• Strategic Approaches: It recommended three core strategies, such as positioning "privacy-first" as a competitive advantage.
• Actionable Play-by-Play: This was the most impressive part. It provided a day-by-day pre-work plan, outlining specific conversations to have with each stakeholder and what questions to ask. It even created a minute-by-minute agenda for the final leadership meeting.
• Back-Pocket Questions: To ensure full preparation, it anticipated curveball questions that might come up in the meeting and suggested how to answer them.

This workflow could be incredibly helpful for any PM. It’s a way to turn the anxiety of a high-stakes problem into a structured, confident plan, which saves hours of prep time and leads to better outcomes.

Conclusion

What I love about Prerna’s workflows is how they demonstrate AI’s ability to solve both highly technical and deeply human problems. The FDA submission tool shows how we can use AI as a co-founder and developer to build real, working applications that deliver huge value and societal benefit. The communication coach shows how we can use AI as a strategic partner to get better at our own core skills, like communication and influence.

What connects both workflows is a pragmatic, problem-first approach. By clearly defining a painful bottleneck, Prerna was able to architect elegant, effective AI solutions. I encourage you to think about the most time-consuming, energy-draining parts of your own work—whether it's wrestling with RegEx or preparing for a tough meeting—and consider how you might build your own AI assistant to help.

Thank You to Our Sponsors

A huge thank you to our sponsors who make this show possible:

• CodeRabbit—Cut code review time and bugs in half. Instantly.
• Lovable—Build apps by simply chatting with AI

Episode Links

• Watch on YouTube
• Listen on Spotify
• Listen on Apple Podcasts