carbon-aware-computing

Distributed LLM pretraining during renewable curtailment windows 🌱

Learn to optimize machine learning tasks for environmental sustainability. Discover how to use real-time electricity data and low-carbon energy sources for model training and inference, reducing the carbon footprint of your cloud operations.

Hackathon winner at AI Engineer World Fair Hackathon: Transforming code, one function at a time, to reduce digital carbon footprints and create a more sustainable digital world.

End-to-end AI deployment decision system combining real model benchmarking, system-level optimization, and infrastructure-aware trade-off analysis (latency, cost, energy, carbon).

GPU-accelerated data-center power forecasting (LSTM/TCN with CUDA mixed precision, quantile P10/P50/P90) with rolling backtests, CPU vs GPU benchmarks, experiment tracking, FastAPI serving and an LP carbon-aware compute scheduler.

Carbon-aware serverless request routing with forecasting, reproducible experiments, and explicit sustainability metrics.

Advisory GPU/HPC workload scheduling research prototype for evaluating energy, carbon, and deadline-aware placement.

Blackout Markets is a shadow optimizer for AI infrastructure teams. It recommends when GPU workloads should run, wait, or move regions based on energy cost, carbon intensity, GPU capacity, latency, reliability, and policy constraints.

EcoLogic is a local Streamlit toolkit for generating and evaluating algorithmic refactors across single files or full codebases. It predicts energy use with feature-based ML, profiles Python/C++/.NET/Java workloads, delivers optimized code with SHAP-powered explainability, and creates shareable PDF certificates for auditable, interpretable results.

Carbon-, cost- and cooling-aware scheduling of AI training jobs across data-center regions. A Pyomo MILP solved with CBC, compared against greedy baselines in a Streamlit dashboard.

Transfer-aware carbon-efficient scheduling framework for simulating distributed deep neural network inference at the edge.

GreenScheduler is a C-based Linux background Daemon that enables Carbon-Aware Task Scheduling by polling RESTful Carbon-Intensity APIs. It dynamically defers Low-Urgency tasks to Periods of Lower Grid Emissions while Executing High-Priority Operations immediately, complete with a Live Analytics Dashboard.

Reproducible artifact for green-energy-aware online multi-objective DAG workflow scheduling.

An auditable carbon-aware VCG mechanism for allocating scarce shared GPU-hours in organizations. We compare exact multi-team VCG allocation with all 720 random-order FCFS outcomes, verifying truthful reporting, externality payments, carbon–access trade-offs, reproducible code, Colab, and a public audit.

Variability-aware low-code DSL that compiles MLOps pipelines to multi-provider IaC/PaC (Terraform, Kubernetes, Argo), with cost, carbon and sovereignty-aware placement and drift-driven adaptation. GetCaaS contribution to the ANR AdaptiveMLOps project (ANR-24-IAS2-0004).

Pre-alpha carbon-aware DevOps / CI/CD reference toolkit.

Carbon-aware cloud workload scheduler that reduces emissions by intelligently shifting workloads across time using multi-objective optimization.

A lightweight pipeline for carbon-aware job scheduling using 72-hour forecasts of grid renewable energy share.

Sustainable AI infrastructure prototypes, starting with carbon-aware workload scheduling

Sub-Watt Saccadic Vision & Power-Aware KV-Cache Engine for Hyperscale Agent Data Centers (70%+ Energy Reduction).