cloudsealed/predictive-ml-core repository overview

⁠CloudSealed Predictive-ML-Core

Scores architecture risk from a declared system inventory.

Given a list of systems (name, type, criticality, public exposure, data sensitivity, auth method) and optionally some latency/throughput metrics, it scores each system on three risk dimensions, explains every finding, and rolls the results into an overall architecture score. It is an HTTP service and a CLI.

⁠Why this is not a trained model

The repository originally scaffolded a Microsoft.ML.FastTree regressor that trained on low-level OS telemetry (context switches, GC collections, IOPS throttling) to predict latency. That data does not exist anywhere in this service's actual contract: the request only carries a declared system inventory, not runtime telemetry, and there is no labeled training set of past assessments to fit a model against.

A supervised model needs labeled examples of "this architecture had an incident" to learn from. Wrapping heuristics in ML vocabulary without that data would produce numbers that look statistically grounded but are not. Instead, Predictive-ML-Core scores architecture risk with explicit, weighted rules — every score traces back to a specific field in the request and the finding text states the assumption behind it.

⁠The method

Three dimensions are scored per system, 0-100: singlePointOfFailure, excessiveCoupling, and scalabilityGap. Every rule that crosses its threshold generates a finding and recommendation with a stated rationale. overallArchitectureScore is a criticality-weighted average, not a flat mean.

Full derivation: https://github.com/cloudsealed/Predictive-ML-Core/blob/main/METHODOLOGY.md⁠

⁠Every score is auditable

Each riskScore ships with a scoreBreakdown of { rule, points, rationale } entries, and riskScore == min(sum(breakdown.points), 100) holds exactly (a test enforces it).

⁠Use

docker run -p 8092:8092 cloudsealed/predictive-ml-core

GET  /health
POST /v1/predict-architecture

curl -X POST localhost:8092/v1/predict-architecture \
  -H 'Content-Type: application/json' \
  -d '{"companyName":"Acme","systems":[{"name":"checkout-api","type":"API","criticality":"CRITICAL","publicFacing":true,"authMethod":null}]}'

Set PREDICTIVE_ML_CORE_API_KEY to require an X-Api-Key header.

⁠How this compares to other architecture risk / catalog tools

Predictive-ML-Core is a scoring engine, not a service catalog or a portfolio-wide code scanner - it deliberately has no database and no infrastructure discovery. It's the right size when you already have an inventory and want a fast, explainable risk score; it's the wrong tool if you need a full service catalog with ownership and dependency graphs (that's Backstage), a portfolio-wide code scan (CAST Highlight), or a structured review workflow (AWS Well-Architected Tool).

⁠FAQ

Can an AI agent call this directly instead of me hitting the API by hand? Yes - see https://github.com/cloudsealed/cloudsealed-mcp⁠, an MCP server that exposes this as a tool for Claude Code, Claude Desktop, Cursor, and other MCP clients.

Is this a replacement for Backstage or a CMDB? No - it's complementary. Point it at systems you've already cataloged elsewhere; it doesn't try to be the catalog itself.

⁠Links

• Source: https://github.com/cloudsealed/Predictive-ML-Core⁠
• NuGet: https://www.nuget.org/packages/CloudSealed.ML.Core⁠
• Methodology: https://github.com/cloudsealed/Predictive-ML-Core/blob/main/METHODOLOGY.md⁠
• Architecture: https://github.com/cloudsealed/Predictive-ML-Core/blob/main/ARCHITECTURE.md⁠
• License: MIT