Chronos

Inspiration

With growing concerns around climate change, rising cloud costs, and strict data regulations, we wanted to build a solution that future-proofs ML pipelines. Chronos was born to empower developers and organizations to retrain models intelligently, by optimizing for time, cost, and carbon emissions—without compromising on performance or compliance.

What it does

Chronos is a carbon-conscious, time-aware AutoML system that:

Enables retraining from past checkpoints with newly tuned hyperparameters.

Supports auto-retraining under updated constraints like CO₂ budgets, cloud cost limits, and regulatory changes.

Integrates with carbon tracking APIs to monitor and minimize emissions.

Dynamically adjusts model architecture, hyperparameters, and compute strategies to stay within sustainability thresholds.

How we built it

Backend: Python + FastAPI for orchestrating training workflows.

ML Pipeline: HuggingFace Transformers + Optuna for hyperparameter tuning.

Checkpoints: Stored and versioned using DVC (Data Version Control).

Carbon Tracking: Integrated with CodeCarbon to estimate and monitor emissions.

Scheduler: Custom module to trigger retraining based on cost/CO₂ drift or policy change.

Frontend: Lightweight dashboard using React to display training stats, carbon usage, and version comparisons.

Challenges we ran into

Balancing performance vs. carbon output in real-time.

Handling checkpoint compatibility between model versions.

Adapting hyperparameter tuning to include carbon and cost as optimization objectives.

Ensuring retraining logic didn't overfit to constraints and degrade model quality.

Accomplishments that we're proud of

Implemented a functional carbon-aware AutoML pipeline.

Successfully retrained models within CO₂ and cost thresholds.

Created a plug-and-play scheduler for constraint-driven retraining.

Developed a time-travel mechanism to roll back and optimize from past checkpoints.

What we learned

How to integrate sustainability goals directly into ML training loops.

Real-world trade-offs between model performance and environmental impact.

The importance of building modular, constraint-aware ML pipelines.

How carbon tracking APIs work and how to integrate them with AutoML workflows.

What's next for Chronos

Integrate with cloud compute APIs to autoscale training based on real-time energy mix (e.g., more training when renewable % is high).

Add explainability tools to justify model changes after retraining.

Open-source the scheduler and retraining engine for broader use.

Expand to multi-cloud setups and federated learning scenarios.

Collaborate with policy and compliance teams to incorporate regulatory frameworks directly into the pipeline.

Built With

• and
• and-automl-engine-mongodb-atlas-?-cloud-database-for-models
• and-gpu-metric-logging-codecarbon-?-real-time-co?-emissions-tracking-and-reporting-ollama-(gemma-2b)-?-local-llm-for-conversational-insights-and-explaining-constraints-?-backend-&-data-management-python-?-core-logic
• and-metadata-?-frontend-react.js-?-interactive-dashboard-and-user-interface-typescript-?-type-safe-frontend-development-chart.js-/-recharts-?-visualizations-for-carbon-usage
• constraints
• cost-trends
• logs
• machine-learning-&-automl-scikit-learn-?-model-training-and-evaluation-(e.g.
• model
• model-versioning
• random-forests)-mlflow-?-experiment-tracking
• training-pipeline