Mounika Maddula

Product Count at Self-Checkout Counter – Fractal Analytics

• Designed and trained a YOLOv8-based object detection model to recognize store products in checkout areas.
• Integrated OCSort for real-time multi-object tracking and applied custom tracking filters to handle occlusions.
• Developed a custom post-processing algorithm to resolve ID switching and maintain accurate product counts.
• Optimized the pipeline for deployment on embedded systems with TensorFlow Lite

Automatic Rooftop Damage Detection – Fractal Analytics

• Built a Mask R-CNN-based segmentation model for detecting damaged rooftops in aerial images.
• Created a multi-step pipeline: initial segmentation, contour-based region refinement, and hierarchical classification.
• Integrated Swin Transformers for advanced feature extraction and fine-grained classification of roof damage types.
• Deployed the trained model via TensorRT for efficient GPU-based inference on large-scale datasets.

Car Scratch and Dent Detection – Fractal Analytics

• Developed a deep learning pipeline using DetectoRS for scratch and dent detection in vehicle images.
• Trained a custom dataset with optimized anchor sizes and aspect ratios for improved accuracy.
• Compared performance of anchor-based vs. anchor-free object detection models to refine precision.
• Implemented model quantization and inference optimizations to enable real-time detection on edge devices.

Table Extraction from Scanned Financial Reports – WNS Global

• Trained Faster R-CNN and YOLO V3 for detecting tabular structures in scanned financial documents.
• Implemented a two-stage pipeline: table detection followed by OCR-based text extraction and cell structuring.
• Developed a Bi-LSTM-based classification model to categorize extracted tables into financial categories.
• Standardized extracted tables into predefined formats using NLP-based row and column mapping techniques.

Automatic Speech Recognition – Quantiphi, Inc.

• Developed an end-to-end speech recognition model using a Convolutional Recurrent Neural Network (C-RNN).
• Extracted MFCC features and trained a hybrid CTC-attention model for robust phoneme recognition.
• Integrated a KenLM statistical language model to enhance decoding accuracy and improve word predictions.
• Optimized model inference with beam search decoding and CPU acceleration techniques.