GreenOps

Agent Flow Diagram
Architecture Diagram
Forecasting Analyser
Optimisation Recommender
Impact Analyser

Inspiration

Cloud waste isn’t just a cost problem—it’s a carbon problem. While working with GCP environments, we discovered that teams often over-provision resources "just to be safe," leading to sky-high bills and unnecessary CO₂ emissions. Worse, optimizing manually is slow and error-prone.

We built GreenOps to automate sustainability. Our vision: an AI team that continuously audits, forecasts, and optimizes cloud infrastructure

What it does

Imagine typing a question like:

“How can I reduce cost and emissions in us-central1?”

GreenOps activates. Here’s what happens:

The GreenOps Agent, our commander-in-chief, kicks into action.
It routes the query to a sequence of expert sub-agents—each trained for a precise role: scout, analyze, recommend, forecast, automate, summarize.
Within moments, you get:
- Detailed infrastructure recommendations
- Forecasted resource + carbon usage
- Execution plan (if it’s safe)
- Weekly report in Google Docs
- Downloadable Presentation deck

🧠 How Each Agent Works

Agent Flow

🔎 optimization_advisor_agent

The strategist. Executes a three-agent relay:

infra_scout_agent – Builds & runs BigQuery SQLs for region data
workload_profiler_agent – Detects idle resources, forecasts emissions with Climatiq
recommender_agent – Crafts a clean recommendation deck

OptimizationRecommendations

📈 forecasting_tool_agent

Predicts CPU, memory, and carbon emissions. Uses BigQuery ML models under the hood and returns formatted results.

ForecastingAgent

⚖️ impact_calculator_agent

"What if we moved from E2 to N2 in europe-west1?" This agent answers using cost and carbon deltas.

ImpactCalculator

🛡️ safe_executor_agent

Before executing changes, it forecasts risks. If safe, it:

Stops instance
Modifies type
Restarts instance Autopilot meets green governance.

📊 summary_generator_agent

Gathers regional data → Runs forecasts → Gets recommendations ➡ Generates beautiful Google Docs reports, including:

Carbon trend charts
CPU vs Carbon overlays
Regional underutilization maps

🖼️ presentation_generator_agent

Auto-summarizes reports and generates slide decks (with visuals!) using python-pptx, uploaded to Google Drive.

🧠 It’s your executive briefing... automated.

Auto Generated Content

Summary Report: Click Here

Presentation Deck: Click Here (Microsoft Office Format, or ONLYOFFICE.app for MAC)

How we built it

We used the Google ADK to build modular, memory-aware agents that pass data between each other like a well-run team. ADK’s SequentialAgent and ToolAgent features allowed us to manage complex flows with simplicity.

Technologies Used:

GreenOps-Architecture-Diagram

Google ADK – Core multi-agent framework
Vertex AI – Smart LLM-driven decision-making (Gemini)
BigQuery – Data storage + forecasting ML models
Climatiq – Emissions API
Cloud Run – Serverless deployment of backend + agents
Streamlit – Frontend interface
Google Docs + Drive APIs – Auto-generated reports and slides
Google Secret Manager – Secure API key storage

Challenges we ran into

True multi-agent logic: Designing agents that not only talk to each other but understand the sequence and context.
Forecast before action: We wanted safety-first automation. That meant building forecast-aware logic for each recommended change.
Dynamic document generation: Building reports and slides that don’t look like they came from a bot.
Chaining LLM outputs: Ensuring agent-to-agent memory is passed correctly, especially in sequential chains with deeply nested outputs.

Accomplishments that we're proud of

Production-ready multi-agent orchestration (not just a demo)
Cross-agent memory, forecasting, and safe execution logic
Climate-conscious AI—bridging DevOps and sustainability
Dynamic document + slide generation with AI
Completely autonomous workflows, from question → insight → execution

What we learned

Good agents = good architecture: Clear responsibilities, tight loops, and minimal memory leaks.
Forecasts are critical: We can’t trust AI to make infra changes without good predictive signals.
Presentation generation with python-pptx is awesome but tricky: Especially when blending LLM text and chart data dynamically.
ADK unlocks real-world agent coordination: It's more than a playground—it's a production-ready toolkit.

What's next for GreenOps

FinOps Agent: Detect budget anomalies and alert teams.
Agent Self-Training: Use past optimization data to refine future decisions.
🧠 New agents:
- security_auditor_agent → for compliance and exposure risks
- cost_anomaly_detector_agent → alert on sudden usage spikes