Interactive 3D pipeline: telemetry, features, AI ensemble, PPO agent, queue, then health surveillance with auto-rollback.
AI right-sizing: t3.micro to t3.nano, $3.80/month saved, with GBDT, Borg safety gate, anomaly and PPO checks all passing.
Explainable decisions: before/after cost and capacity, multi-model safety radar, and the telemetry features driving each call.
30-day cloud spend forecast with 95% confidence bounds. Forecaster R² 0.930, selectable 7 to 90 day horizons.
Verified model metrics: right-sizing 99.67% accuracy, forecaster R² 0.930, anomaly AUC 0.8416, PPO agent 150 episodes.
Cloud connect: least-privilege read-only IAM policy and one-click CloudFormation template, with AWS, Azure and GCP tabs.
Decoupled queue dispatch: signed optimization payloads go to AWS SQS FIFO, Azure Service Bus or GCP Pub/Sub, deduplicated.
Rollback state machine: dispatched, 600s surveillance, then auto-rollback if CPU exceeds 85% or commit. Spend forecaster below.
Queue dispatch panel: action, environment, risk tier and the raw JSON message contract, behind an operator approval gate.
Closed-loop flow: pick cloud, collect metrics, AI models, risk-gated action, queue, resize, health monitor, then rollback or commit.

Inspiration

Cloud bills quietly balloon. Teams over-provision VMs "just to be safe", idle workloads keep running, and nobody has time to manually check every instance across AWS, Azure and GCP. Existing tools usually stop at a dashboard and leave the risky part, actually changing production infrastructure, to a human. We wanted a system that can recommend, explain, execute and undo its own changes.

What it does

NeuroSwarmOpti is a multi-cloud fleet optimization platform.

Ingests live telemetry from AWS CloudWatch, Azure Monitor and Google Cloud Monitoring.
Right-sizes resources: a model ensemble classifies each VM as scale_down, scale_up or no_action, picks a target SKU and predicts the dollar savings.
Scores anomaly risk (0-100) so risky changes are flagged for approval, while low-risk ones can be automated.
Predicts workload resilience using a deep network trained on Google Borg cluster traces.
Forecasts spend over 7, 30 and 90 days and warns about budget overruns.
Explains every decision with plain-language reasons and SHAP feature attribution.
Self-heals: after every change there is a 10-minute observation window. If CPU saturates above 85%, it automatically rolls back to the exact previous state. Otherwise it commits the change.

How we built it

Models: LightGBM classifier and regressor for right-sizing, Isolation Forest plus supervised LightGBM for anomaly detection, a PyTorch residual network (BatchNorm, Dropout, cosine annealing) for workload resilience, gradient boosting for cost forecasting, and a PPO reinforcement-learning agent for scaling policy.
Data: four datasets, including a multi-cloud VM dataset, 277k anomaly records, 1.3M Google Borg trace rows and 54k cloud billing records.
Backend: Python with a FastAPI REST server, plus a Typer CLI.
Cloud integration: boto3, Azure SDKs and Google Cloud SDKs, with job dispatch over AWS SQS, GCP Pub/Sub and Azure Storage Queue.
Frontend: a monochrome web console with Chart.js charts and a 3D workflow view.
Safety: least-privilege IAM/RBAC policies, an approval gate for risky actions, and captured original_state snapshots for rollback.

Challenges we ran into

Normalizing three very different cloud APIs, each with its own metrics, instance types and queue systems, into one schema.
Making automation safe. Resizing a live VM can hurt performance, so we built the approval gate and the observation-and-rollback loop.
Anomaly detection is a heavily imbalanced problem, and we had to tune for high recall so real risks are not missed.
Keeping the models fast enough to give recommendations in real time.

Accomplishments that we're proud of

Right-sizing accuracy of about 99.5% on the multi-cloud dataset.
Cost forecasting with an R² of about 0.93.
A full closed loop: collect, predict, explain, queue, execute, monitor, then commit or roll back.
Support for AWS, Azure and GCP behind a single interface.

What we learned

Explainability is what makes people trust automation.
A reliable undo path matters more than a clever model.
Real infrastructure telemetry is messy, and preprocessing took as much effort as modeling.