George
Debrah

Research Vision

Global trade and logistics networks are among the most complex optimization environments in existence — thousands of interdependent decisions made under uncertainty, across time zones, regulatory boundaries, and fluctuating demand. Yet most systems still rely on static rules and reactive planning. My research asks: what becomes possible when ML and AI move from describing logistics problems to actively solving them?

My undergraduate thesis — combining YOLOv5 object detection with Bee Colony Optimization to cut vehicle wait times by 32% in a dynamic traffic environment — is the clearest expression of this conviction. Traffic management and freight routing share the same mathematical DNA: both are real-time, multi-agent, combinatorially explosive problems where learned models and adaptive algorithms together outperform any fixed strategy. The thesis was a proof of concept. Logistics is the application at scale.

My long-term goal is to develop and deploy ML/AI systems inside real logistics and trade operations — in industry R&D settings where research translates directly into measurable throughput, cost reduction, and resilience improvements.

Existing Research: Intelligent Traffic Optimization

My undergraduate thesis at Ashesi University produced a fully implemented intelligent traffic management system for Ghanaian urban intersections — one of the first studies to combine deep learning-based vehicle detection with a nature-inspired optimization algorithm in a sub-Saharan African road context.

I trained a YOLOv5 object detection model on 113,888 annotated traffic light images, achieving 75.8% mean average precision (mAP) on validation data. The detection model fed real-time vehicle counts into a Bee Colony Optimization (BCO) algorithm that dynamically adjusted signal phase timings across intersections — replacing fixed, pre-timed cycles with an adaptive system that responds to actual traffic load. The BCO solver achieved a lower optimal fitness value than baseline static schedules, corresponding to a 32% reduction in average vehicle wait time and a 28% improvement in overall traffic flow efficiency in simulation, validated against real-world traffic datasets from Ghanaian intersections.

This work established three principles that now shape my broader research agenda: that perception and optimization must be co-designed rather than bolted together; that nature-inspired search algorithms are particularly well-suited to dynamic, high-dimensional scheduling problems; and that domain-specific datasets — not generic benchmarks — are essential for building systems that actually work in context. These principles translate directly into how I approach ML/AI applications for trade and logistics.

ML-Driven Route & Network Optimization

Classical vehicle routing and freight planning algorithms — TSP variants, column generation, branch-and-price — are powerful but brittle: they optimize for the world as modeled, not the world as it unfolds. Disruptions, port delays, carrier failures, and demand spikes break static plans immediately. My research explores how reinforcement learning and predictive ML can be integrated with optimization solvers to produce plans that are not just optimal at dispatch time, but dynamically re-optimized as conditions change.

Directly extending my thesis work, I am interested in how population-based metaheuristics — which proved highly effective for adaptive traffic signal phasing — can be adapted for multi-depot, multi-commodity freight routing under stochastic demand and real-time disruption signals. The goal is a planning system that learns from historical shipment data to anticipate bottlenecks before they materialize.

Predictive Analytics for Supply Chain Resilience

Supply chains fail quietly before they fail visibly — demand signals drift, supplier lead times shift, and inventory buffers erode gradually until a stockout or shipment crisis makes the problem undeniable. My experience building an ML anomaly detection model achieving 91% precision at Pensive Global Consult, and constructing forecasting pipelines at TheBoardroom Africa that improved prediction accuracy by 35%, established a strong foundation for applying these methods at the supply chain level.

I am interested in early-warning ML systems that continuously monitor supply chain health — tracking lead time variance, supplier reliability scores, port congestion indicators, and macroeconomic signals — and surface actionable alerts before disruptions propagate downstream. The critical research challenge is building models that remain calibrated as the distribution of disruption events shifts, which is precisely the adaptive learning problem my thesis work began to address.

Intelligent Data Pipelines for Trade Operations

Logistics organizations generate enormous volumes of operational data — from warehouse management systems, IoT sensors, ERP platforms, customs documentation, and carrier APIs — that sit in silos, arrive late, or require manual reconciliation before they can be used. My production experience building optimized data pipelines at Phrontlyne Technologies, reducing retrieval latency by 1.8 seconds and reporting time by 50%, and deploying ERP systems configured for IFRS 17 compliance, gave me direct insight into how data infrastructure quality determines the ceiling of any analytical or AI capability built on top of it.

I am interested in the design of real-time data integration architectures for trade operations — systems that ingest, clean, and route operational signals fast enough for ML models to act on them, not just analyze them after the fact. The research question is not just "what can we predict?" but "can we deliver that prediction in time for it to change a decision?"