Projects
OpenCV, PP-OCRv5, PyTorch
-
Developed a custom text extraction pipeline for Indian ID
documents using OpenCV by fine-tuning a PP-OCRv5 model (~90%
accuracy).
-
Delivered structured JSON data for 5,000+ documents to enterprise
endpoints.
-
Optimized OCR pipeline for CPU-only inference, cutting
infrastructure costs by 40%.
DeBERTa-v3, Wave2Vec2, PyTorch
-
Developed a speech-based grammar scoring engine to predict scores
(1-5) from 45-60s audio samples using MFCC, spectral, and prosodic
features and text embeddings.
-
Fine-tuned DeBERTa-v3-large for regression over 409 training
samples.
-
Achieved an RMSE of ~0.42 on validation set through regularized
regression head and hyperparameter tuning.
PyTorch, Word2Vec, BERT Fine-tuning
-
Built a custom PyTorch model with Word2Vec and prototyped BERT to
identify semantically similar question pairs with 84% accuracy
(0.82 F1).
-
Trained a Word2Vec-embedded neural network using Bayesian
hyperparameter tuning, achieving 12% accuracy gain over baseline.
- Preprocessed 400,000+ question pairs for training.