Feb – May 2026 · Complete

TextScope — NLP Pipeline & Model Optimization

Solo build

AI/ML

How NLP models actually work From tuning to meaning

Why I built it

I wanted to understand how NLP models actually work — how text becomes something a model can use, how training happens, and why one approach beats another. So I built the pipeline by hand rather than calling a library that would have done it in three lines and taught me nothing. Every stage is one I wrote myself specifically so I would have to understand it. Partway through I started writing the process up on Substack, which turned out to be the part that exposed what I only thought I understood.

What I built

Built an end-to-end NLP pipeline processing raw text through tokenization, bigram extraction, and TF-IDF feature engineering — each stage written out rather than imported, so nothing was a black box.
Trained and evaluated logistic regression models across L1 and L2 regularization to analyse feature importance and model performance.
Designed the experimentation workflow that compares results across hyperparameter configurations, so runs are reproducible and comparable rather than one-off.
Structured the codebase around swappable wrappers — separate TF-IDF, embedding, and evaluation modules behind a common interface — which is what made the representation comparison a change of one component rather than a rewrite.
Built an interactive Streamlit app on top of the pipeline for exploring how models behave on new text.
Wrote two beginner-facing articles explaining what I learned, working from the pipeline itself rather than from other explanations of it.

What I had to learn

How the whole path from raw text to a prediction actually fits together, at the level of detail you only get from having written each stage yourself.
That representation dominated model choice on this dataset. Swapping TF-IDF for sentence embeddings moved cross-validated accuracy far more than any amount of tuning the classifier did.
Cross-validation as a discipline rather than a step — comparing configurations honestly means fixing everything else first, which is a harder engineering problem than it sounds.
That writing an explanation is a good test of whether you have one. Several things I thought I understood turned out to be gaps only when I tried to put them in a paragraph for someone who did not already know the answer.

Results

Switching from TF-IDF to sentence embeddings raised cross-validated accuracy from 0.50 to 0.96.
Two published write-ups explaining how NLP models work and how to improve them, aimed at people starting from zero.

Stack

Pythonscikit-learnsentence-transformersPyTorchpandasNumPyStreamlit