Setting up my ML environment from scratch: MedMind

#python #ai #productivity #beginners

I decided to build a clinical AI system from scratch. Not using the OpenAI API. Not a tutorial project. An actual system where I train my own model, build my own search pipeline, and deploy the whole thing.

This post is about Day 1 — setting up everything before writing a single line of ML code.

Why I'm doing this

Most AI projects I see from students are just API wrappers. You call GPT-4, get an answer, show it in a UI. That's fine, but it doesn't teach you how any of it actually works.

I wanted to understand what happens under the hood — how models get trained, how retrieval works, how you actually serve a model in production. So I picked a real problem (clinical decision support) and decided to build the whole thing myself.

What I'm building

MedMind — a system that takes a clinical question, searches a database of medical knowledge, and generates an answer using a model I trained on real medical exam questions.

The full stack:

Download and clean a real medical dataset
Fine-tune a language model on that data
Build a RAG pipeline with a vector database
Evaluate the model honestly
Serve it with FastAPI
Build a UI with Streamlit

Setting up the environment

First thing — Python version matters a lot in ML. I installed Python 3.11 specifically because PyTorch and HuggingFace have best support for it.

Then I created a virtual environment. This keeps all the libraries for this project separate from everything else on my machine:

python -m venv venv
venv\Scripts\activate  # Windows

Then installed the core libraries:

pip install torch transformers datasets peft trl accelerate
pip install chromadb sentence-transformers
pip install fastapi uvicorn streamlit

Each of these does something specific:

transformers — gives access to pre-trained models like OPT, Mistral, LLaMA
peft — lets you fine-tune models efficiently using LoRA
trl — makes instruction fine-tuning easier
chromadb — the vector database for storing medical knowledge
sentence-transformers — converts text to vectors for search

Project structure

I organized everything before writing code:

medmind/
├── data/       ← data scripts
├── training/   ← fine-tuning
├── rag/        ← retrieval pipeline
├── eval/       ← evaluation
├── api/        ← FastAPI backend
└── frontend/   ← Streamlit UI

For training — Colab

My PC has no GPU. Training a language model on CPU would take weeks. So I set up Google Colab with a free T4 GPU. This is what most people without expensive hardware do — it's completely normal.

DEV Community