The architecture

One LLM, routing to many small models.

Medico doesn't try to be one giant model that knows everything. It's an orchestrator: a language model that reads your case and calls the right specialist model for the job — most of them already trained and shared by researchers on Hugging Face and Kaggle.

Hugging FaceMelanoma MobileNetScreenerDermatologyHugging FacePneumoniaMobileNetV3 TinyRadiologyHugging FaceDiabetic RetinopathyMobileNetV2OphthalmologyHugging FaceMalaria CellMobileNetV2PathologyKaggleWound TypeClassifierWound Care / PodiatryKaggleGeneral Skin DiseaseMobileNetDermatologyMedico+ORCHESTRATOR
Calling Melanoma MobileNet Screener · Hugging Face — Dermatology

Section A

The idea

Training a single model that diagnoses skin cancer and reads chest X-rays and interprets ECGs from scratch would take an enormous dataset, a lot of compute, and months of work — and it would still be mediocre at each task.

So we don't. For almost every one of these problems, someone has already trained a small, focused model and published it — on Hugging Face, Kaggle, or as a public model or dataset page. These models are tiny (often a few megabytes to a few hundred MB), they run fast, and they're genuinely good at their one job.

Medico's job is orchestration: a language model reads what you typed or uploaded, figures out which specialist model(s) are relevant, calls them as tools, and then explains the combined result in plain language. The LLM is the coordinator — the small models are the experts.

Section B

Why small, specialized models

01

Don't reinvent the wheel

Reuse models and datasets the community already published for focused clinical image tasks, such as HAM10000, Chest X-Ray Pneumonia, and APTOS.

02

Fast and cheap

Available entries prioritize focused CNNs, small transformers, OCR, and rules workflows. GPU-required entries are marked unavailable in the catalog.

03

Clear availability

Each entry is labeled as available or not available so the catalog does not imply every model can run in the current app.

04

Easy to swap & audit

Each model does one narrow task, so you can benchmark it, replace it, or add a new specialist without touching the rest.

Section C

What's surprising

You'd be surprised how much signal these little workflows pull out of ordinary inputs — a phone photo, a report scan, or one exported imaging slice:

Image

A compact vision model can screen a skin lesion from a phone photo or diabetic retinopathy from a fundus image.

Report

For ECG, blood, and lab reports, OCR plus explicit threshold rules can be more practical than a heavy image model.

Scan

Chest X-rays, dental X-rays, and MRI slices can route to separate model entries instead of one generic classifier.

These models are for screening and decision-support, not diagnosis. Always confirm findings with a qualified medical professional.

Section D

End to end

  1. 01

    You describe a case or upload an image, report, or scan.

  2. 02

    The orchestrator LLM decides which specialist models are relevant — it skips the ones that aren't.

  3. 03

    It calls each available model or rules workflow as a tool. GPU-required entries stay marked unavailable.

  4. 04

    The LLM reads the results, cross-checks them, and writes a plain-language summary back to you.