I vibe coded an AI caregiving system for my aging parents. Now I'm building a startup to share the tech with others.

• Srdjan Stakic vibe-coded a security camera system for his parents to ensure their safety.
• Stakic used vibe-coding platform Lovable to get started, as well as popular AI chatbots.
• His vibe-coded software became the basis for his AI-assisted startup Alvis.

This as-told-to essay is based on a conversation with Srdjan Stakic, 49, a former film producer who vibe-coded an AI system to monitor his elderly parents and detect falls. He's now launching a company that aims to offer the technology to others. This interview has been edited for length and clarity.

When I was diagnosed with stage four cancer two years ago, AI became essential.

Everything was happening so quickly: The doctors would talk to me for 15 minutes and leave me with more questions than answers. AI gave me an objective way to document and make sense of what was happening.

I'm now in remission. As my health improved, my parents' health declined, and I began helping them with cooking, cleaning, and medical appointments.

English was not their first language, and communicating with healthcare providers was tough. I recorded our conversations with their doctors and compared them to the after-visit summary using AI. I would put together all this information and translate it into Serbian for them.

But I soon wanted more than what the chatbots could offer. I wanted a system that could observe what was happening with my parents, or any other patient, and assess it through the lens of safety and dignity. I would think of how much guilt I'd feel if something happened and I wasn't there. What kind of son would I be?

I had never coded before, and I didn't have millions for an initial investment

I don't have a background in coding. I have a doctorate in health education and a master's in film production, and I have produced some films of my own.

I started outlining my idea with Gemini and ChatGPT to examine it from a tech and ethical standpoint. I built this document of what I wanted to achieve. I kept asking my family how they wanted to be treated in each scenario — like a fall or medical emergency — and I wanted to make the system flexible.

Then I transferred to Lovable. Lovable gave me a live development environment where I could describe what I wanted, see it built in real time, test it, and iterate. It connected the pieces, the frontend, the backend, the database, the authentication, the integrations, things I did not even know I needed until they were there. The chatbots helped me plan. Lovable helped me build.

I uploaded hundreds of training videos for nurses and healthcare providers to train the AI. I created a high-fidelity validation pipeline and a labeled dataset. I labeled real-world caregiving footage with established clinical benchmarks, like Stanford's C-I-CARE framework. When you approach a patient and introduce yourself, you tell them why they're there, you ask the patient's name and pronouns, and you introduce what you're about to do. You explain next steps and see if they have any questions or concerns.

I also started building an AI equipped with cameras to identify falls. I would fall in the middle of my living room and see whether the system recognized that and how long it would take.

It took me a few months to make it work

I tried different cameras and protocols, but ultimately, I had to hire an IT company to help me connect multiple cameras. The system can now identify a fall and send notifications to loved ones or EMS, and provide their location with a brief summary of their health records. The system also analyzes interactions between caregivers and my parents. It's sophisticated enough to analyze in real time — based on audio and video — if a caregiver is being rude or unprofessional. My parents have felt safer since I built this. I also built a feature that scans their environment for any trip hazards, such as cables.

I don't want to spy on my family, so I don't actively review all the video footage. When a concern is flagged, the system clips approximately 30 seconds around that moment and notifies mewith a summary of what it observed and why. It can also generate an advocacy letter from that same analysis: what was said, what was done, and how the interaction compares to the C-I-CARE framework to evaluate caregiver conduct.

I launched a company to offer this tech to others

This all started as an idea for my family, but the more I talk about it, the more people tell me they wish they had this for their parents. So I decided to launch a startup, called Alvis, to make this system available to others.

It detects falls in real time, recognizes when a caregiver goes above and beyond, and generates advocacy letters when something goes wrong. It's in private beta and accepting waitlist applications for our pilot cohort, launching April 13. The model will be a monthly subscription, similar to what families already pay for camera cloud storage, with a premium tier for AI-assisted analytics.

This week, my mom was hospitalized, and I used AI in four ways

First, I used it as a real-time medical interpreter: Every lab result went straight into Claude, so I understood what was happening immediately, not the next morning when a doctor was free.

AI was also my clinical advocate. When a history and physical exam understated her cancer history, Claude caught it. When her glucose started climbing from steroids, Claude flagged it.

Third, I used AI to translate updates into patient-friendly language in both English and Serbian.

Finally, Alvis — the camera system I designed — was running live in her hospital room all night, with her permission and a nod from her care team. It picked up her saying in Serbian, quietly, that she had endured too much. It flagged when I visited, and we recorded ourselves together.

It's amazing to see how vibe coding is democratizing access to AI tools. You can build a company that helps a very niche group that needs a specific thing. I still don't fully understand code or the extent of what I built, but it seems to be working.