Skip to content
Imtiaz Hossain

the playground / 4 experiments

Every experiment here runs on authentic outputs from models I trained, with the exact metrics from the result files. Segment a tumor, probe a text detector, or trace 27 experiments in a single grid. Nothing is fabricated, and the architecture is built so a live inference endpoint can be swapped in later without changing a single screen.

exp-01Computer Vision

Brain Tumor Segmentation & Classification

Segment tumors in T1-weighted MRI slices and classify them across four categories. Compare U-Net against Attention U-Net and inspect real inference panels from 860 unseen test samples.

test dice

88.22%

exp-02NLP

AI-Generated Text Detection

Detect ChatGPT-written text from 28 handcrafted stylometric features. Explore feature importance, the HC3 benchmark, and why detection collapses across generators.

hc3 accuracy

97.40%

exp-03NLP

News Topic Classification

Nine architectures times three preprocessing pipelines on 102k headlines. See where BERT wins, where a 19-second Bi-GRU nearly matches it, and why stemming hurts WordPiece.

best macro-f1

0.9376

exp-04Machine Learning

Software Quality Prediction

An honest study of a hard problem: predicting quality labels from nine code metrics. Explore the cluster structure and why the signal is weak, reported without varnish.

weighted f1

0.4418