AI Systems Engineering Portfolio

From Brainwaves to Brain Scans: A Robust Neural Network for EEG-to-fMRI Synthesis

https://github.com/kgr20/E2fNet

Autonomous Investment Intelligence Agent

The Investor Intelligence Agent is a comprehensive tool for investment analysis, providing a user-friendly platform for gathering and analyzing investment-related information.

The system architecture features a streamlined interface that allows users to access information through a webpage dashboard, a chatbot, or downloadable CSV files. The platform is built on a technical stack that includes Streamlit for frontend development and Python for backend processes. The backend components include an NLP engine for chatbot interaction, a Neuro-Fuzzy Model for financial risk assessment, and web scraping tools such as TagUI and yfinance for data extraction. Data is stored securely in a SQLite database.

The platform includes several key features:

Company Background & Financial Analysis: The system extracts key company information and financial metrics using TagUI, Beautiful Soup, and yFinance Python packages. Technical analysis is performed using TagUI and OpenCV to capture and interpret stock price movements.
Competitor Analysis: Competitor data, including ratings and reviews, is collected using TagUI for benchmarking purposes. Future improvements could include incorporating financial metrics, technical analysis, news sentiment analysis, and ESG scores.
ESG Integration: The platform integrates environmental, social, and governance (ESG) factors with financial metrics. It uses fine-tuned BERT models (EnvRoBERTa, SocialBERT, GovRoBERTa) for analyzing ESG data. The system architecture is modular, with components for environmental, social, and governance data analysis. Real-time data processing is supported by API integration, and the platform offers flexible scoring approaches.
Real-Time News Analysis: The platform includes a module that gathers and analyzes real-time news using API and RPA technologies. Sentiment analysis is conducted using NLTK’s VADER lexicon, and duplicate removal is handled through TF-IDF vectorization and cosine similarity.
Neuro-Fuzzy Recommender System: The platform employs an Adaptive Neuro-Fuzzy Inference System (ANFIS) to predict stock risk based on recent stock data. The system includes layers for fuzzification, rule application, and defuzzification to process and output predictions.

Deep Learning-Based Geospatial Analytics Platform

The project focused on developing a comprehensive system for monitoring and classifying forest cover using satellite imagery. The primary goal was to leverage advanced image classification models to assess forest health and track changes in land cover over time, particularly in detecting deforestation and other environmental shifts.

The system architecture included three deep learning models—PSPNet, U-Net, and Fully Convolutional Network (FCN)—optimized for semantic segmentation tasks. PSPNet utilized a Pyramid Scene Parsing Network to capture contextual information at various scales, achieving a validation accuracy of 86.15%. U-Net, with its U-shaped architecture, was the most effective, reaching a validation accuracy of 90.03%. The FCN model, configured with 14 input channels, delivered a validation accuracy of 87.51%. An ensemble method combining the outputs of these models was explored, but U-Net was ultimately selected for its superior performance.

Time-series analysis was integrated into the system to track changes in forest cover over different periods. By comparing satellite images across time points, the model provided quantitative and visual insights into land cover changes, aiding in the detection of environmental trends.

A service platform was developed, featuring a web-based interface that allows users to upload dual satellite images for processing. The backend, built with FastAPI and TensorFlow, handled the classification tasks and generated detailed reports on land cover changes. The platform's design emphasized ease of use, enabling users to interact with and interpret the data efficiently.

Intelligent Innovation Strategy Analytics Platform

Developed an intelligent agent to perform comprehensive landscape scans of companies, value propositions, and cross-sector offerings, expected to generate €5+ million per year. The project received endorsements and sponsorship from McKinsey's Head of Digital, Vancouver Office Managing Director, Head of Research, and other senior partners. It has been deployed with several leading financial institutions in Europe and Latin America.

Automated Business Report Parsing and Data Extraction System for Consultants

Semantic Force is a platform developed to streamline the analysis of complex business documents, including annual reports, 10-K filings, and relevant news articles. The platform is built to cater to the needs of consultants and analysts who require quick and accurate insights from large volumes of text data. At its core, SemanticForce employs advanced Natural Language Processing (NLP) models, such as T5 and BART, to automatically generate concise summaries of financial documents. These summaries focus on extracting key financial metrics, strategic insights, and sentiment analysis, offering a comprehensive overview of a company's financial health and market position. The platform also incorporates a sentiment analysis module to assess media and public perception, which can be particularly valuable in understanding the broader context of a company's operations. One of the practical features of SemanticForce is its Q&A system, powered by the LLaMA-2 model. This allows users to query the processed data directly, extracting specific information without the need to manually sift through entire documents. The system is designed to handle a range of queries, making it a useful tool for consultants who need to access relevant data efficiently.

From a technical perspective, the platform integrates Python with NLP libraries like Huggingface and Langchain, and uses a Retrieval Augmented Generation (RAG) framework to enhance its information retrieval capabilities. This allows SemanticForce to process and analyze both structured and unstructured data, making it adaptable to various types of inquiries. The backend architecture supports secure data management, with documents stored in PDF, CSV, and Pickle formats, and all API calls are secured through token-based authentication. The frontend is built with Streamlit, providing an intuitive interface where users can upload documents, submit queries, and receive results in a straightforward manner. Feedback from users is collected to continuously improve the system’s performance, ensuring that the platform evolves to meet the needs of its users effectively.

The platform was developed with the practical needs of the consulting industry in mind, offering a reliable and efficient solution for analyzing large datasets. While still in its initial stages, the platform demonstrates potential for significantly improving the way consultants and analysts work with complex financial information. The focus is on providing a tool that integrates seamlessly into existing workflows, delivering actionable insights with minimal manual effort.