change-order-analysis

What it does

Implements machine learning-based analysis for construction change orders, including predicting costs, classifying change order types, identifying patterns, and streamlining approval workflows.

How it works

• Provides a comprehensive system with ML components: a ChangeOrderPredictor that uses TF-IDF text features plus numeric features to classify change order types and predict cost and schedule impacts. It includes training on historical data, transforming descriptions into features, and fitting classifiers/predictors (RandomForestClassifier for type, GradientBoostingRegressor for cost and schedule). It stores an encoder for type labels and a text vectorizer.
• Includes a ChangeOrderManager to create, update, and summarize change orders, track costs and schedule impacts, and compute a simple severity based on cost and schedule relative to contract value.
• Offers a ChangeOrderAnalyzer to convert change orders into a DataFrame and analyze by type, including counts, total and average costs, and total schedule days.
• Exposes a pattern for a ChangeOrderClassification workflow, a CostBreakdown, ScheduleImpact, and ChangeOrderDetail data structures to manage attributes such as co_id, co_number, title, description, justification, co_type, status, and attachments.

When to use it

• Use the ML classifier/predictor after you have historical data with fields: description, co_type, cost_impact, schedule_impact, contract_value, project_phase, and affected_elements_count. Use train() to fit models, then predict() to get predicted type, costs, and schedule for new descriptions.
• Use the ChangeOrderManager to create and manage CO lifecycle, update costs and schedules, and generate summaries for reporting.
• Use the pattern analysis to explore COS by type and other attributes to inform decision-making and identify dominant drivers.

What it can touch

• Tools: claude-code
• It interacts with Python code, data structures, and ML models (RandomForestClassifier, GradientBoostingRegressor, TfidfVectorizer). It relies on historical_data as a pandas DataFrame with specified columns during training, and uses numpy and joblib for model persistence.

Caveats

• Models require training data; performance depends on historical quality and feature engineering. Training uses a specific feature setup: TF-IDF features with max_features=500 and 1-2-gram ranges, plus contract_value and affected_elements_count as numeric features.
• Classification and prediction outputs include probabilities and costs/schedule estimates, but actual outcomes depend on inputs and model fit.
• The workflow assumes defined enums and data classes as shown (ChangeOrderType, ChangeOrderStatus, ImpactSeverity), and relies on consistent field names in historical data.