November 5, 2025. LiveScope

LiveScope is an internal predictive-estimation tool developed to support DeltaV Live conversion projects. During the early stages of a project, engineers may spend several weeks reviewing the existing control system and estimating the labor required to complete the work. These estimates directly affect Control Southern’s ability to prepare accurate, competitive bids.

LiveScope streamlines this process by converting structured control-system data into a predicted engineering-hours estimate. What began as a series of spreadsheet-based experiments evolved into a web application that uses historical project data to produce consistent, data-driven estimates. The tool supplements an engineer’s judgment while reducing the time required to develop an initial project scope.

The project’s primary technical challenge was transforming inconsistent plant exports into a standardized feature set. Raw customer files rarely matched the structure used to train the models, and different control-system elements often used different column names and workbook layouts. To address this, I developed a data-preparation pipeline that filters elements marked “In Scope,” identifies the relevant worksheet, and maps plant-specific columns into a shared LiveScope schema.

Each mapping is validated against known reference workbooks before the data is passed to the model. Seemingly minor inconsistencies, such as whether an unused feature is represented by a blank cell or a zero, could materially change calculated totals. Resolving these differences required flexible column matching and validation rules rather than fixed spreadsheet positions.

LiveScope maintains two modeling approaches. The “no reduction” model predicts an initial estimate before accounting for repeated or highly similar system elements. The “reduced” model is trained on engineers’ final adjusted estimates, which reflect the productivity gains created by duplication. Maintaining both models allows engineers to compare an unconstrained first-pass estimate with a more practical projection of the work likely to be required.

Both models were built using RandomForestRegressor pipelines. Hyperparameters were tuned with RandomizedSearchCV and five-fold cross-validation before the selected models were evaluated on an 80/20 holdout set. Once validated, each model was retrained on all available labeled data for production use. When the feature set changed, the models and their associated metadata were regenerated together to ensure that production inputs remained aligned with the training schema.

The strongest results came from the no-reduction model, which achieved a holdout R^2 of approximately 0.88 and a mean absolute error of approximately 0.85 hours. These results indicate that the model explained most of the observed variation while maintaining a relatively small average prediction error on unseen data.

LiveScope demonstrates how data standardization, supervised machine learning, and practical software development can be combined to turn inconsistent plant exports into a repeatable estimation workflow. By shortening the initial estimating process and providing a consistent analytical benchmark, the tool helps engineers prepare faster, more defensible project bids.