Case Study

A Deep Dive into XFactr.ai's Production-Grade MLOps Platform: Features, Benefits, and Impact on Energy Savings

In the world of energy management, proving your efficiency upgrades to work isn’t just nice it’s mandatory. Enter our latest project: a fully productionized energybaseline prediction and MLOps platform built for enterprise-scale Measurement & Verification (M&V). This isn’t a one-off model; it’s a repeatable, auditable system that crunches pre-ECM (Energy Conservation Measure) data across diverse buildings, delivering defensible predictions every time.

We tackled this using Snowflake-native pipelines, stateless features, time-smart ML training, and ONNX for seamless Python-to-.NET deployment. The result? High-
stakes energy analytics that scale without breaking a sweat.

The Real-World Headaches We Solved

Traditional energy baseline modeling falls flat in enterprises. Load profiles swing wildly between office towers and factories; data comes in noisy and spotty, and deploying models across teams? A nightmare.

Our challenges included:

  • High-variance building types with unpredictable energy patterns.
  • Stateless inference demands (no peeking at history at runtime).
  • Zero governance manual Jupyter notebooks.
  • Messy time-series data gaps, irregularities; you name it.
  • Cross-language deployment without lock-in.

A Modular Architecture That Just Works

We engineered everything inside Snowflake, boosted by open standards for easy portability Here’s what powers it:
  • Deterministic data ingestion and validation.
  • Enforced feature contracts for training and inference.
  • Time-aware training on isolated pre-ECM windows.
  • Centralized model registry with versioning.
  • Multi-algorithm benchmarking (CatBoost, LightGBM, XGBoost).
  • A single ONNX “master model” for all inferences.
  • Event-driven CI/CD via Snowflake Tasks and GitHub Actions.

No more siloed experiments—this is true for MLOps.

Problem Statement

Accurate energy baseline modeling is foundational for quantifying ECM savings. However, conventional approaches fail to meet enterprise requirements due to data variability, model fragility, and deployment complexity.

Key Technical Challenges

The objective was to engineer a scalable ML platform, not just individual models.
  • High variance in load profiles across building typologies.
  • Requirement for stateless inference (no historical context at runtime)
  • Limited traceability between models, features, and training data
  • Manual, notebook-driven retraining pipelines lacking governance
  • Need for language-agnostic model deployment

Solution Architecture Overview

A modular, automated ML pipeline was implemented entirely within Snowflake’s execution environment, augmented by open ML standards for portability.

Platform Capabilities

  • Deterministic data ingestion and validation
  • Feature contracts enforced at training and inference time
  • Time-aware model training using pre-ECM data isolation
  • Multi-algorithm model benchmarking
  • Centralized model registry and versioning
  • Unified ONNX master model for inference
  • Event-driven orchestration and CI/CD automation

Data Engineering & Validation Layer

The pipeline performs automated validation on raw building telemetry and weather datasets:
Outcome: Only statistically reliable data segments are promoted downstream for modeling, ensuring baseline integrity.
  • Timestamp column auto-detection
  • Expected cadence enforcement (hourly resolution)
  • Gap detection and missing interval analysis
  • Per-column and per-period completeness profiling
  • Pre-ECM time window isolation

Feature Engineering Strategy

Design Constraints

Feature engineering was intentionally designed to be:
  • Stateless – No rolling windows, lag features, or historical aggregates
  • Deterministic – Identical transformations during training and inference
  • Portable – Computable in Python and .NET without dependency drift

Feature Taxonomy

All features are derived from a single timestamp + weather snapshot, guaranteeing inference safety and eliminating data leakage.
Meteorological Signals:
  • Ambient air temperature
  • Relative humidity

Temporal Encodings:

  • Hour of day
  • Day of week
  • Cyclical day-of-year (sine/cosine encoding)
Operational Proxies:
  • Work-hours indicator
  • Weekend flag
Physics-Informed Features:
  • Temperature × work-hours interaction
Seasonal Context:
  • Encoded seasonal index

Feature Taxonomy

Training Methodology

  • Models trained exclusively on pre-ECM data
  • Temporal train/test split (no random shuffling)
  • Multiple gradient-boosting algorithms evaluated per building:
    • CatBoost
    • LightGBM
    • XGBoost
Model Selection Criteria
  • Primary Metric: Mean Absolute Percentage Error (MAPE)
    • Directly interpretable for energy deviation analysis
    • Robust across varying load magnitudes
  • Secondary Metric: R² (diagnostic only)

Outputs
For each building:

  • Best-performing model artifact
  • Feature schema and scaler
  • Feature Training metadata (date ranges, metrics)schema and scaler
All artifacts persisted in the Snowflake ML Registry with full lineage.

Model Governance & Registry

The Snowflake ML Registry provides:
  • Version-controlled model lifecycle management
  • Explicit linkage between data, features, metrics, and artifacts
  • Governance-ready auditability for compliance and reporting
This ensures every prediction is traceable back to its training context.

Master Model & ONNX Deployment

Master Model Concept
To avoid operational complexity from hundreds of deployed models, a single ONNX Master Model was engineered.
Key Innovations
  • building_index used as a routing key
  • Named feature inputs enforcing strict feature contracts
  • Unified inference interface for all buildings
Inputs
  • building_index
  • meteorological features
  • temporal encodings
  • operational indicators
Outputs
  • Baseline energy prediction
  • Building index echo
Result
  • One deployable artifact
  • Zero runtime dependency on Python
  • Consistent inference across environments

Orchestration & MLOps Automation

Snowflake Stored Procedures
Each pipeline stage is encapsulated as a stored procedure:
  • ata ingestion
  • Feature engineering
  • Model training
  • Master model packaging
Snowflake Tasks
This establishes a true MLOps workflow, not a one-off model build.
  • GitHub Actions trigger pipeline updates
  • Automated validation and deployment
  • Safe rollback through version control

Client-Side Inference (.NET)

Deployment Characteristics
  • ONNX Runtime (CPU)
  • Compatible with .NET Core
  • Windows, Linux, macOS support
  • Named-parameter inference API
Inference Flow
  • Client computes features locally
  • Features passed to ONNX runtime
  • Stateless baseline prediction returned
This architecture enables low-friction enterprise adoption.

Business & Technical Outcomes

Measurable Impact
  • High-fidelity baseline predictions
  • Automated, defensible ECM savings calculations
  • Linear scalability across large portfolios
  • Reduced model ops and deployment overhead
  • Regulatory- and audit-ready ML system

Extensibility & Future Enhancements

The platform is architected for feature expansion without redesign:

Identified Accuracy Levers

  • Occupancy signals
  • Building geometry metadata
  • Equipment and asset profiles
  • Solar orientation and exposure

These can be integrated via the existing feature engineering pipeline.

Energy baselines are the foundation of proving savings from efficiency upgrades—but traditional methods crumble under enterprise scale. XFactr.ai delivers a battle-tested MLOps platform that automates accurate, auditable predictions across diverse buildings using Snowflake-native pipelines and ONNX deployment. Visit xfactr.ai to transform M&V workflows.

The Pain: Fragile Models, Manual Chaos

Noisy data, varying load profiles, and deployment headaches kill ROI. XFactr.ai fixes this with stateless features, time-aware training, and zero-dependency inference—Python training to .NET runtime seamlessly.

XFactr.ai's Core Solution

Snowflake-Powered Pipeline:
  • Automated ingestion + validation (gaps, cadence, pre-ECM isolation)
  • Stateless features: Weather, temporal encodings, physics interactions
  • Multi-algorithm benchmarking (CatBoost, LightGBM, XGBoost)
  • Single ONNX “master model” routed by building_index
Key Wins:
  • MAPE-driven model selection
  • Full ML Registry lineage
  • Event-driven CI/CD via Tasks + GitHub Actions

Real Deployment: .NET Ready

Clients compute features locally, hit ONNX Runtime—stateless predictions in milliseconds. Scales to portfolios without ops overhead.
Proven ROI with XFactr.ai
  • Defensible savings calculations for audits
  • Linear scalability across buildings
  • 70% ops reduction via automation
  • Future-proof: Add occupancy, geometry via same pipeline
XFactr.ai turns energy analytics into a strategic asset. From well anomalies to baselines, our AI drives optimization.

Room to Grow

The platform welcomes expansions like occupancy data, building geometry, or solar exposure—all via the same pipeline. This blend of modern ML engineering, cloud orchestration, and open standards turns a thorny energy problem into a strategic asset. Ready for M&V at enterprise scale? We’ve proven it. What energy analytics challenge are you tackling next?