Nwalor Chibuzor Ignatius

I built a lightweight LLM-based graph context builder that learns to traverse knowledge graphs from a query and return relevant nodes, with the goal of replacing heuristic GraphRAG retrieval with a learned traversal policy.

I developed a training pipeline using Qwen3, Hugging Face, TRL, and PEFT/LoRA, starting with supervised fine-tuning on synthetic graph-traversal data and progressing to reinforcement learning for retrieval quality and traversal efficiency.

I designed benchmarks against conventional GraphRAG retrieval across retrieval quality, traversal efficiency, latency, and inference cost, with experiments spanning model size and inference configurations.

An AI-powered application that helps developers break down and understand complex codebases by explaining subsystems within an app at various levels of abstraction, as well as providing the framework for tools like automated code reviews, large-scale refactors, etc.

I was part of the team that developed autonomous soccer-playing robots. I specialized in image processing and object detection with OpenCV and the robot's self-localization. Technologies used were Python, ROS, and OpenCV.

A lightweight hybrid CNN-LSTM architecture for performing command word recognition. I handled designing and implementing the architecture as well as optimizing the model in terms of model size and speed of the whole system (preprocessing and prediction).

I developed a mobile app to help solar energy users track their energy consumption and analyze the usage data. This project was built with Ionic and powered by a Django back end.

I developed a mobile app that allowed users to browse through and preview media and then download after making a payment. This was built with Ionic and Django.

I built a performant gallery web app with infinite scrolling with Angular 2 for a hackathon.

I wrote a Python wrapper for the Paystack.com API.

I used Three.js to make a 3D preview to allow users see how different clothes would look on a model.