I vibe coded an AI tool to help my mom fight stage 4 cancer. It helped us catch errors in her treatment and let her die with dignity.

• Pratik Desai, 34, vibe coded a platform to manage his mother's cancer care, improving her treatment.
• Desai later refined the workflow with advanced coding tools; it caught errors in her care plan.
• Desai stressed that though his app isn't perfect, it helped his mom die with dignity.

This as-told-to essay is based on a conversation with New Jersey-based Pratik Desai, 34, who vibe coded a tool — which he later refined with advanced coding tools — to help his mother navigate Stage 4 duodenal adenocarcinoma, an advanced cancer affecting the small intestine. His words have been edited for length and clarity.

I don't really have a medical background; I've just been healthcare adjacent throughout my life because my family's in it.

After graduating, I took a job at Accenture and focused on systems integration, business analysis, and the crossroads of how technology impacts workflows, mainly working on medical and non-medical benefits at the government level. I moved to a boutique strategy consulting role where my major client was Johnson & Johnson. I also built an app for my wife to help her remember to keep her medical credentials up-to-date.

I found my way to Salesforce, which was my first real dive into AI. I worked my way up to be the global practice lead for personalization, which was synonymous with machine learning. I started my own AI company called 1to1, which exited to a company called ListEngage, where I now sit on the leadership team.

I also launched a podcast to expose people to AI and what it can do for them, whether it be caregiving, their health, or anything else in their life. While AI is giving people more information than ever before, there's still a gap in people's understanding of how to use it.

My mom was a completely healthy woman

On October 20 of last year, she threw a 70-person Diwali party. I'd never have guessed that exactly a month later, she'd be diagnosed with stage 4 duodenal adenocarcinoma. We got the diagnosis after she had some stomach issues and we pushed her to go to the hospital.

The doctors' plan was to discharge her without an oncology appointment, which we felt made it clear that there was very little time for her and that the doctors were writing her off. We basically were given this diagnosis and a message of "good luck."

I'm a type A person, so I wasn't going to take that as an answer. I'd never been a caregiver before. I don't know what it looks like without AI. I took the tools in front of me and asked, what can I do now?

We decided we were going to fight — and AI would help us

I first asked AI to make me an expert on stage 4 duodenal adenocarcinoma to find out what we were up against. AI began coaching us and synthesizing what the doctors were saying.

Once we decided to fight, I used AI to do deep research on every hospital in New Jersey. I called every number I could, trying to schedule something within the week, even though it was Thanksgiving.

It was AI that found, on a hospital's oncology department web page, a doctor who happened to be my ex-girlfriend from high school. She got back to us and got us an appointment.

My mother lived for 76 days, 67 of which were inpatient. There was not a single instance where we felt like the medical world was actually looking out for her goals as a patient — mainly, an optimized length of life to say the goodbyes she wanted to say. Instead, it felt like they were always seeing how fast they could get us to the next step.

At one point we were making critical decisions based on a CAT scan. The workflow identified there were two misdiagnoses in the CAT scan report, and three instances in which the wrong cancer was stated in the report. Neither my mother nor I would've been able to read that report, but AI was able to read it and probe deeper.

The workflow I ended up creating that really worked over time is a daily export from the Epic system. We put that into NotebookLM, along with any symptoms that I saw or anything she told me. I'd say to NotebookLM, "synthesize the data."

Then I'd go to my preferred LLM, which was Claude, and ask it to explain what the information meant. I'd ask things like, "What should I know about for tomorrow's appointment and the scenarios that are in front of us?" Later, I'd go back and ask, "What are the second opinions that I should ask for? What doesn't look right here that I should be pushing back on?"

The workflow wasn't perfect to start

What we built is still not perfect or elegant, but it's accessible and free.

I wanted to keep things as simple as possible, but as her medical records grew to 1,600 pages long, not including any of the images or scans, we couldn't load it into a single thread anymore. Eventually it was just way too much data, and I ran out of context.

I first used Google AI Studio, which worked fine, but I ended up moving away over time because Claude's models were just getting better. I wanted to use the best possible model, knowing I had such a short amount of time left with my mom.

I was by her bedside from 5 a.m. to 10 p.m. every single day throughout the 76 days she was alive, and NotebookLM was my second opinion and coach.

Some doctors I've talked to about my product have pushed back and said that AI can hallucinate and is only right 70% of the time. I responded, what if we took the medical system and graded it the same way? We expect AI to be perfect, but the medical world isn't.

I believe the AI saved her life three times

There were no less than three times that, guided by this workflow, we caught emergency situations and I believe saved her life.

On Christmas day, I noticed something kind of off in the way she was walking, breathing, and talking. She was pushing back on spending time with the family. The medical system was not returning a call on Christmas.

I typed what I was seeing into AI, and it determined she was absolutely dealing with complications of a pulmonary embolism. I didn't immediately rush to the hospital, but it allowed me to send an SOS to my cousin, who is a doctor, and he said to get her to the hospital. The emergency line may have taken us four or five hours to get a call back.

AI was also able to detect a pattern of her starting to bleed out seven days after her blood transfusions. We realized that 48 hours after an infusion, they'd take her off of a liquid-only diet and move her to a food diet, which would irritate the ulcer that she had and expedite bleeding. The amount of blood that she had lost would be so critical that she would essentially be dying. We caught this twice.

AI moved the needle for my mother

We were able to ask the right questions and push her doctors to do better. By doing so, we were able to give her the time necessary to say her goodbyes. She was able to kiss my daughter, who is two.

I'm of the age where a lot of my friends have family members in hospital systems, and I've enabled a few of them with this kind of workflow. One of my buddies felt like his mother's medical team was not doing the right thing, so he used this workflow, studied up, and was able to stay on top of her care team.

In one instance, he called a meeting with her doctors, and he said they were thoroughly impressed at his understanding of the entire case without a single note in front of him. They kept asking if he was a doctor or biology major, and he said he's a marketer. That's the power of the workflow.