Olusanmi Hundogan

I built a multi-agent AI chatbot with specialized agents for communication (Google Calendar and Slack), finance (Yahoo Finance), and more. I also leveraged the A2A protocol for agent-to-agent coordination and MCP for seamless tool integration, enabling the bot to perform complex automation workflows end-to-end.

I developed a back-end service for extracting named entities—persons, organizations, and locations—from news text. The system leverages a fine-tuned DistilBERT model that I adapted for the domain using pseudo-labelled data generated via GPT-4, resulting in a solution that outperforms other approaches in precision and flexibility.

The architecture follows a modular, extensible design pattern inspired by Scikit-learn, allowing easy integration of multiple model types. I implemented and evaluated various extractors, including rule-based methods, SpaCy pipelines, transformer-based models, and LLMs—each wrapped in a unified interface for training, prediction, and benchmarking.

The final model is exposed via a high-performance REST API built with FastAPI and deployed using Docker. The solution is scalable, containerized, and ready for production, with the back end capable of reliably and efficiently serving real-time inference requests. This project spanned the full machine learning lifecycle, including data preprocessing, chunking, label generation, fine-tuning, evaluation, and deployment. It was a robust foundation for advanced named-entity recognition (NER) use cases in news and media processing.

I developed a scalable machine learning framework to predict corporate CO₂ emissions across Scope 1, 2, and 3 using heterogeneous financial, operational, and sustainability data.

Designed modular pipelines per scope with configurable preprocessing, feature selection, imputation, and segmented ensemble regression (EvenWeightMiniModelArmyEstimator) to address sparse, skewed targets. I conducted extensive experimentation with transformations, voting strategies, and bucket classification, achieving low sMAPE (< 7%) and well-calibrated confidence intervals on a held-out dataset of 8,000+ companies. The system, deployed via FastAPI on DigitalOcean, supports single and bulk predictions and imputes missing emissions data, enabling regulatory compliance and enhanced ESG reporting. The solution demonstrates end-to-end capability from data integration and model optimization to deployment and horizontal scaling under concurrent requests