Kingsley Nwabueze Jide

• Designed a scalable BIM agent architecture for IFC knowledge-base querying, achieving sub-second response times across large structured datasets.
• Improved retrieval relevance with a hybrid search layer combining ChromaDB vector search with Neo4j graph queries.
• Built a WebRTC voice pipeline (STT / TTS / VAD) on LiveKit, holding low-latency audio control at >1300ms round trip.
• Shipped an IFC.js-based 3D visualization layer for interactive BIM model exploration and spatial understanding of building components.

• Built a high-performance ingestion pipeline in Python processing up to 166 records per minute.
• Engineered a real-time scoring and classification layer in C++ and Python for downstream lead qualification.
• Automated cold-calling with a voice agent workflow orchestrated in n8n.
• Applied feature engineering and validation to raise dataset reliability for analytics consumers.

• Built a Discord monitoring extension for bot detection using ML and real-time streaming analytics.
• Integrated webhook-driven security alerts, cutting response time by 40%.
• Shipped responsive front-end interfaces with a tightly coupled back-end for live alerting.
• Tuned model evaluation to minimize false positives and lift detection accuracy.

The project produced a three‑axis cutting‑force dynamometer that integrates a T‑shaped mild‑steel shank, a carbide cutting insert, and a Wheatstone‑bridge strain-gauge network; raw bridge voltages are read by an Arduino microcontroller running custom C++ firmware that implements the calibrated conversion equations for tangential, feed, and radial forces, performs real-time computation of force values, and writes the results to an SD-card data logger; the firmware was validated against a simulated finite-element calibration (showing 0.1–17% error) and against linear and non-linear predictive models, achieving R² values between 0.80 and 0.93, thereby confirming that the embedded system accurately captures and stores the experimental cutting-force data.

This research addresses a major bottleneck in deploying deep learning models to resource-constrained edge devices (like smartphones and IoT hardware) by optimizing model pruning—the process of cutting redundant neural connections to make networks smaller and faster. Traditional methods rely on simulated annealing (a physics-inspired trial-and-error search) regulated by a "blind," preset cooling schedule that frequently freezes too early or wastes computation. To solve this, this work develops a closed-loop hybrid controller using a Proportional-Integral-Derivative (PID) feedback loop (the same math behind vehicle cruise control) to act as a smart thermostat. By dynamically monitoring the search speed and adjusting temperature in real-time, the controller prevents optimization stagnation and ensures a highly efficient search path.