Edge-AI-Triage

Inspiration

I noticed the clinic only ran for a couple hours each day while studying for board exams. The doctor had to head back to the city after that so people just waited around. It got me thinking about how hard it is for some places to get even basic advice when the real help is not around much. Software might give some idea of what to do next with symptoms but only if it works where phones and clinics are both scarce. I am not sure how far that can go though. I talked it over with a friend one afternoon at the park. Turns out a lot of people do not trust anything AI says about health. Others do not even have a phone to run it on. Most would rather just sit and wait for an actual doctor. That led me to wonder how these tools handle different skin tones or groups of people. Representative data seems important and so does proper testing instead of hoping it works the same everywhere. Questions like that pushed me to start something called Edge AI Triage. It is just a prototype right now looking at local software for early checks on symptoms. Model size power needs cost and how steady the whole thing runs all matter a lot. Some of it still feels unclear to me.

What it does

Edge AI Triage is a prototype exploring how locally running software and AI-related technologies could help people understand possible next steps when they experience health symptoms. The idea came from observing people in my local area waiting to consult a doctor at a clinic that operated for approximately two hours a day because the doctor also had to travel to the main city. I began wondering whether affordable technology could help bridge part of the gap between experiencing a health concern and understanding what to do next. The project is still a research-oriented prototype, not a clinically validated diagnostic system. It is not intended to replace qualified healthcare professionals, and its reliability and response behaviour require further testing.

How we built it

I built this thing called Edge AI Triage with Python handling most of the backend stuff. There were some frontend parts that talked to it too. It also had image processing and model inference mixed in. The idea was to see how all of it might run together as a local system that could help with healthcare stuff.

Getting back into real Python work after three months off was rough at first. Python 3.14 gave me a bunch of problems so I switched over to 3.12 instead. That switch took maybe an hour and it was a reminder that picking the right setup matters.

Once that was sorted the frontend and backend started talking to each other. The whole thing still needs more debugging and testing though so the outputs actually change when the inputs do. I am not totally sure how far the improvements need to go yet. It seems like responses are still too similar across different cases.

Challenges we ran into

Working with limited resources on my laptop was tough from the start. The development environment and different parts of the project all had to run together on it which made everything slower and more annoying. Sometimes it felt like the laptop was just not into it.

The AI side was tricky too. I had trouble picking a good base model and the right data especially for something health related. These models need more than just producing an answer. Their behavior has to be checked against real evidence and that part still feels messy to me.

I also wondered how skin tones and different populations might change the results. Finding data that actually represents people fairly and setting up a proper evaluation is something I still have to figure out better.

Knowing seven languages got me thinking about how symptom descriptions can shift depending on background. Translating the words is not always enough because context and common expressions matter a lot. That makes building something accurate way more complicated than just adding languages.

The most annoying issue was the system giving basically the same response across different cases. It seemed fine at first but then I started wondering if it was really processing inputs or just stuck on one output. That reminded me a demonstration does not always mean the logic is sound.

These problems pushed me to question my assumptions more and check things step by step. I am not totally sure I caught everything yet.

Accomplishments that we're proud of

I am proud of taking something I noticed locally and actually building it into software instead of leaving it as just an idea. I started looking into the technical parts that could make it work and that felt like progress even if it was slow at first.

Recognizing the identical response problem was another part I feel good about. I did not treat the first demo as proof that everything was fine. Finding the flaw was disappointing but it gave a clearer sense of what to check next.

Beyond the code the project has also got me asking about healthcare accessibility offline computing language differences data representation and responsible AI. I am not totally sure how all of those connect yet.

What we learned

This project showed me that making AI stuff is way harder than just getting a demo to look good. You can have something that talks to a backend and shows answers but the real logic might still be off.

I learned a lot about testing different things and seeing how the data flows. Also managing what depends on what and making sure something really works instead of just seeming to. But the biggest thing for me was realizing that even when things move slow these projects can still be worth it.

There were times the problems felt too big especially with the hardware limits and getting parts to connect. But when I showed what I was doing to others and they got excited it gave me a push to keep going. That did not solve the issues but it helped me find the energy again to try more.

It also made me see that this kind of project involves people just as much as the tech. Like how they talk or what they believe or if they can actually use it. I am not sure how all that fits together yet.

What's next for Edge-AI-Triage

I want to make edge ai triage more practical by trying out a physical device with a raspberry pi. It would be affordable and standalone so there is less need for a phone each time someone uses it.

Placing it in a community setting with a simple interface could help. Some people do not have smartphones or do not feel comfortable with apps that are too complicated.

The raspberry pi might run light software locally. This would cut down on needing internet all the time.

I think that sounds good but the design depends on a few things. Hardware performance is one and model size is another along with power use and cost. Software reliability matters too. Maybe it will all work out but I am not totally sure how it fits together yet.

Built With

• computer-vision
• deep-learning
• edge-ai
• explainable-ai
• healthcare
• image-classification
• machine-learning
• medical-ai
• medical-imaging
• python
• pytorch
• resnet18
• skin-lesion-classification
• triage