Dinesh Chowdary Attota

Credit Model Research - Factor Modeling

• Research on implementing a factor model integrating macroeconomic indicators and time series data to predict consumer credit risk, optimizing model performance for different asset portfolios (e.g., Bank Card, Mortgage).
• Engineered robust feature selection processes using advanced statistical techniques, improving model accuracy and reducing overfitting in the context of large-scale economic datasets.
• Implemented scalable linear models using PySpark and proprietary bureau tools to manage, preprocess, and model large volumes of consumer data efficiently.
• Collaborated with cross-functional teams to refine model architecture, ensuring alignment with evolving business requirements.
• Conducted deep analysis of static and time-series data for different consumer credit portfolios, testing multiple model configurations to identify optimal predictive features and temporal dynamics.
• Executed complex SQL queries to extract and manipulate vast consumer datasets, optimizing data retrieval and preparation for in-depth statistical analysis and model development.

Pertinent skills:Factor Modeling & Predictive Analytics, Feature Engineering & Selection, Big Data Processing & Distributed Computing, SQL & Data Query Optimization, Stakeholder Collaboration & Cross-functional Communication

LLM Research - Global AI Agents (ongoing research)

• Leveraging Large Language Models (LLMs) to generate user-defined functions (UDFs) for bureau-specific globals, streamlining code generation with efficient prompting techniques.
• Developing a Graph-based Retrieval-Augmented Generation (RAG) system using Neo4j to encode global specifications, integrating it as a dynamic knowledge base for LLM-enhanced query generation.
• Utilizing in-context learning and few-shot learning strategies to generate JavaScript-based UDFs and their corresponding SQL queries, optimizing for precision and relevance.
• Fine-tuning a specialized LLM model to efficiently reduce globals generation time by 80%, significantly outperforming traditional programming approaches.

Natural Language Processing Research:

• Conducted in-depth research on behavioral health classification using sensitive police narrative data, leveraging deep learning techniques such as transformer-based models to propose novel approaches for improving classification accuracy. Utilized statistical analysis methods to rigorously evaluate model performance.
• Led the implementation of advanced Automatic Speech Recognition (ASR) models, using transformer architectures to convert speech data into text. Employed noise reduction and signal processing techniques to improve the clarity and accuracy of transcriptions.
• Modified transformer architectures for enhanced model performance, adding auxiliary information and embeddings mid-training to enrich contextual understanding and improve model adaptability to sensitive text data.
• Developed and trained models for speech diarization, enabling the differentiation of multiple speakers within audio recordings.
• Utilized Python-based tools (Pandas, NumPy) and visualization libraries (Matplotlib, Seaborn) for statistical analysis and performance monitoring, designing experiments in Jupyter Notebooks to ensure the validity, reliability, and interpretability of experimental results.
• Presented research findings at academic conferences, workshops, and seminars, synthesizing complex technical concepts into accessible presentations, papers, and posters for both technical and non-technical audiences.
• Collaborated closely with faculty members and research teams to design and execute experiments, ensuring robust data collection, result analysis, and peer-reviewed publication of research findings.

LLM Research:

• Implemented a robust framework using Large Language Models (LLMs) for image-captioning tasks, generating detailed captions for input images. Designed a looping process where captions are iteratively refined and used to regenerate images as similar as possible to the original using state-of-the-art diffusion models.
• Leveraged the LLama Instruct and other similar models to generate detailed, context-driven questions based on initial captions.
• Applied reinforcement learning techniques to fine-tune LLM models based on feedback from the quality of generated outputs.
• Developed a novel metric for evaluating image similarity, incorporating object detection and image segmentation rather than relying solely on traditional cosine similarity from vision transformer embeddings.

Internet of Things Research:

• Applied deep learning techniques to classify various types of IoT network attacks, utilizing advanced neural networks for anomaly detection and attack categorization across multiple IoT devices and protocols.
• Proposed and implemented novel ensemble learning approaches, combining multiple classifiers to improve the accuracy and robustness of detecting and classifying different types of IoT network attacks.
• Experimented with various autoencoder architectures, including variational and sparse autoencoders, to effectively reduce data dimensionality.
• Published research findings in an IEEE journal, contributing to advancements in IoT security