Work
What I have built, measured and taught.
Each case study says what the question was, what I did, and what came out, with the code where it is public.
Case studies
-
When the document is wrong
Open language models repeat a contradicting document 49–68% of the time. Letting them “think” first costs 10–19× the GPU energy and does not make them resist.
-
HVB-Persian
A fine-tuned ParsBERT classifier fused with VADER lexicon scores reaches F1 of 0.88, 0.94 and 0.95 on Digikala, Snappfood and the combined set of about 84,000 Persian reviews.
-
Tumor segmentation in CT
A 3D U-Net for liver and tumor, and a 2D U-Net for lung tumors trained with oversampling against extreme class imbalance, reaching a validation Dice of 0.80 on lung.
-
The AI Path
A free, open course in six planned seasons, from Python to AI agents. Every lesson follows one rhythm — plan, predict, run, check, explain.
More projects
Smaller builds. Each links to its repository.
-
Retrieval QA over NASA documents (on GitHub)
Question answering over NASA documents with LangChain. Instructor-XL embeddings, a Chroma vector store and Dolly v2 (3B), with answers grounded in the retrieved passages and shown with their sources.
-
Live Bicep Counter (on GitHub)
Counts biceps curls from a webcam in real time, using MediaPipe pose estimation and OpenCV, with an on-screen rep count and progress bar.
-
Persian sentiment with FastText (on GitHub)
A three-class FastText classifier for Persian text (positive, negative, neutral) with 86% accuracy.
-
Classical sentiment baselines (on GitHub)
Naive Bayes, Linear SVC, SGD and Random Forest on TF-IDF features, compared on accuracy, precision, recall and F1. All four reach 0.857–0.862 accuracy.
-
PDF Similarity Comparison (on GitHub)
A small desktop tool that extracts the text of two PDFs and compares them with TF-IDF vectors and cosine similarity.
-
Image Processor (on GitHub)
A Tkinter application for basic image processing, with grayscale conversion, histogram display, histogram equalization and edge detection.
-
Movie Recommendation and Analysis (on GitHub)
An analysis of a movie database, and film recommendations built through feature engineering.