John Dickson

In the context of designing energy-efficient green buildings, stakeholders needed to understand how architectural design choices—such as window size, door size, building orientation, and exterior color—impact key performance metrics like energy consumption, thermal comfort, and building lifespan.
There was no live operational data available since the buildings had not yet been constructed. However, extensive computer simulation (computer experiments) data existed from building energy models.

My task was to quantify the impact of design parameters on building performance and determine the optimal configuration for a house of a given square footage that would minimize overall energy consumption.

I built a surrogate (meta) model to approximate the complex and computationally expensive simulation outputs, enabling fast evaluation of design alternatives. I formulated a Mixed Integer Linear Programming (MILP) optimization model to minimize predicted energy consumption. Finally, I evaluated trade-offs between design options to ensure solutions were practical, interpretable, and architect-friendly

The company was experiencing high customer churn driven by service cancellation calls. Incoming calls were routed using static rules, without considering regional churn patterns, call center performance, or capacity constraints. Historical data showed that certain regional centers were consistently more successful at retaining customers, but this insight was not operationalized in real-time call routing.

I was responsible for designing a data-driven call routing strategy to increase customer retention by directing high-risk cancellation calls to call centers with the highest historical save rates, while respecting operational constraints such as call volume, regional demand, and center capacity.

I developed heuristic-based routing rules as a baseline solution to quickly test impact (e.g., region-to-center affinity, historical save-rate weighting). I designed a Mixed Integer Linear Programming (MILP) optimization model to maximize customer retention probability, and compared heuristic and MILP approaches to balance solution quality, scalability, and operational feasibility. This resulted in achieving a 12% improvement in customer retention, directly reducing churn from service cancellation calls.

Industrial boilers were experiencing unplanned shutdowns, resulting in production downtime and service escalations. Boiler health was manually monitored by operators, who lacked predictive visibility into when a failure would occur. Historical IoT sensor data existed, but it was not being leveraged for proactive maintenance or early warning.

I reduced unscheduled boiler shutdowns by developing a predictive maintenance solution that detected failures early enough for operators to take preventive action. I analyzed historical failure events and high-frequency IoT time-series sensor data collected from boilers in production environments. I built and evaluated multiple models, including LSTM networks for temporal dependency learning and tree-based algorithms for interpretability and robustness. I also designed and deployed an automated alerting pipeline that triggered SMS notifications to boiler operators around 45 minutes before predicted failure, allowing time for corrective intervention.

This successfully predicted 6 out of 10 unscheduled shutdown events in advance, providing operators with actionable lead time. It also reduced total shutdown duration by 60% by enabling proactive maintenance instead of a reactive response.