Clarios Agent-AI-thon 2026
ChillerSense
Multi-Agent AI for Industrial Chiller Plants
Best Idea Award
For the most complex and ambitious agent concept with strong long-term potential.
Mad Scientist Award
For bold, unconventional architecture pushing the boundaries.
The only team to win multiple awards.
Built on real operating data from MEBCO
Primer
What is a chiller plant?
Chiller plants are the cooling backbone of large facilities. Inside, four giant compressors cool production lines, server rooms, and offices.
If a chiller fails, you don't just lose comfort — you lose production hours. Downtime is measured in hours of lost production.
Heat moves from the building to the atmosphere in one closed loop.
The Problem
Huge data, no time to read it
0TR
Plant capacity
0
Sensors generating data 24/7
Operators rely on manual logs and reactive maintenance. Early-warning signals already exist, but buried in the huge amount of data, and no one has time to read them.
By the time an incident is logged, downtime has already happened. The opportunity: turn raw sensor data into proactive intelligence, before the failure, not after.
Architecture
End-to-end pipeline
Two layers, one closed loop. Sensor data flows up; agent intelligence flows back to the operator's inbox.
Data Pipeline
Chillers system
5-sec · 349 sensors
The plant streams raw sensor data every 5 seconds across all four chillers, pumps, and cooling towers.
DAT → Parquet
Columnar storage
Proprietary Chillers System exports converted into Parquet for fast, columnar analytical reads.
Anomaly ML
3 models, 1 vote
Three-model ensemble: Isolation Forest, One-Class SVM, and an LSTM autoencoder. At least 2 of 3 must agree.
JSON Outputs
Structured findings
Every hour produces a JSON record: anomaly score, contributing sensors, severity, and timestamps.
JSON → DOCX
Python converter
Findings are formatted into operator-friendly DOCX briefs that an agent can cite verbatim.
Power Automate
Desktop
Power Automate Desktop watches the local folder and uploads new briefs to the SharePoint knowledge source.
SharePoint
Knowledge source
Single source of truth — all four agents read from the same library, so reasoning stays grounded in the same evidence.
Agent Layer
User Query
Orchestrator
Sonnet
Routes the query, selects specialists, composes the final answer.
Diagnostician
Opus
Root cause + early warning. Cites specific sensors, timestamps, and ML evidence.
Optimizer
Opus
Efficiency + cost. Looks for setpoint drift, load imbalance, and idle compressors.
Reporter
Sonnet
Composes the daily plant intelligence email — clear, dense, no fluff.
Power Automate
Cloud Flow
Outlook + Teams
Delivery
ML Ensemble
Three models, one vote
01
Isolation Forest
Outlier detection
02
One-Class SVM
Boundary learning
03
LSTM Autoencoder
Temporal patterns
Ensemble Vote
≥ 2 of 3 must agree
Anomaly flagged
Each model has different inductive biases. Isolation Forest is fast and global. One-Class SVM learns a tight normal boundary. The LSTM autoencoder catches sequence-level drift. Consensus catches real failures while filtering false alarms.
Agent Loop
Four agents at work
Orchestrator
SonnetHUBRoutes the query, selects specialists, composes the final answer.
Reads
User query · agent outputs
Produces
Final response · daily report
Diagnostician
OpusRoot cause + early warning. Cites specific sensors, timestamps, and ML evidence.
Reads
SharePoint findings
Produces
Evidence-cited diagnosis
Optimizer
OpusEfficiency + cost. Looks for setpoint drift, load imbalance, and idle compressors.
Reads
SharePoint findings
Produces
Setpoint + load recommendations
Reporter
SonnetComposes the daily plant intelligence email — clear, dense, no fluff.
Reads
Diagnostician + Optimizer
Produces
Daily email body
Field Result
Caught 24 hours early
How ChillerSense flagged a chiller failure a full day before operations noticed it.
Day before
NOV 17
Normal day
Anomaly score 0.35
Plant looks normal. Nothing flagged.
Warning day
NOV 18
Anomaly detected
Score 0.44, peak 0.63 at 09:00
6 warning hours. All 3 ML models agreed. Operators had not noticed yet.
Failure day
NOV 19
Failure confirmed
Anomaly score 0.35
Ops logs Chiller 4 pressure relief valve stuck and Chiller 3 decoupled.
Detection lead time
24 hours
When operators reported the failure on Nov 19, ChillerSense had already raised the alarm a full day earlier, with all three ML models in agreement on one of the most anomalous hours of the entire month.
Tech Stack
What's under the hood
Data & ML
Python
Pandas
Scikit-learn
PyTorch
Parquet
Agents & Orchestration
Copilot Studio
Claude
Power Automate
Delivery & Storage
SharePoint
Outlook
Teams