This repository contains applied machine learning projects built with PyTorch and scikit-learn.
The projects demonstrate end-to-end workflows: data preparation, model training, evaluation, fairness checks, and production-ready exports.
Predicts engagement likelihood from text synopses and tabular metadata such as genre, maturity rating, and duration.
- BiGRU with additive attention for text
- MLP encoder for tabular features
- Fusion head with dropout and layer normalization
- Metrics: AUC, Accuracy, Brier Score, ECE (calibration), per-genre AUC
- Exports: TorchScript and ONNX
Folder: multimodal/
Script: multimodal_ctr.py
Implements a Vision Transformer (ViT) from scratch for image classification.
- Patch embeddings, class token, positional encoding
- Multi-head self-attention encoder blocks
- Data augmentation with AutoAugment and RandomErasing
- Metrics: Top-1 Accuracy, Confusion Matrix
- Exports: TorchScript
- Includes attention rollout visualization
Folder: vision_transformer/
Script: vit_cifar10.py
Applies machine learning to optimize the hiring funnel using synthetic ATS/HRIS-style data.
- Candidate success prediction with calibrated classification
- Fairness audit by group with threshold tuning
- Time-to-hire regression with gradient boosting
- Sourcing channel effectiveness and lift analysis
- Funnel anomaly detection with Isolation Forest
- Exports: joblib models and model card
Folder: talent_analytics/
Script: talent_analytics_sklearn.py
- Multimodal Engagement: AUC ~0.87, well-calibrated (ECE ~0.03)
- Vision Transformer: CIFAR-10 Top-1 Accuracy ~80%
- Talent Analytics: Candidate Success AUC ~0.85, Time-to-Hire MAE ~5 days
- Python 3.10+
- PyTorch for deep learning
- scikit-learn for classical machine learning
- TorchScript and ONNX for model export
- joblib for persistence
- Calibration metrics and fairness reporting