CloudSealed.ML.Core 0.2.1

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. A real model becomes viable once enough real assessments accumulate to serve as training data, keeping the same contract.

The method

Three dimensions are scored per system, 0–100:

singlePointOfFailure — base weight by criticality (LOW=0, MEDIUM=15, HIGH=35, CRITICAL=55) plus a modifier by type (DATABASE +15: state is more expensive to replicate; THIRD_PARTY_SERVICE +20: outside your control, no fallback declared; APPLICATION/API +5). The request schema has no redundancy field, so the finding text states the assumption explicitly: single instance, worst case.

excessiveCoupling — proxy for exposure and dependency, since the request carries no dependency graph. publicFacing without authMethod scores highest (+40); with an authMethod declared, less (+15); internal systems get a small base (+5). THIRD_PARTY_SERVICE type adds +25. Fan-out across the whole request — more than two THIRD_PARTY_SERVICE entries — adds an organization-level coupling bonus, capped, since that pattern is a system-wide signal, not just a per-system one.

scalabilityGap — prefers real data: if historicalMetrics is present, p99LatencyMs above 1000ms and a p99/avg ratio above 3 (heavy tail under load) both add weight. Without historicalMetrics, it falls back to a weaker, explicitly conditional signal — DATABASE with dataSensitivity declared, or CRITICAL systems with no load data at all ("unknown risk", not a measurement).

Each rule that crosses its threshold generates the corresponding finding and recommendation — severity is derived from the score, not picked by hand.

overallArchitectureScore is a criticality-weighted average across systems, not a flat mean. A flat mean lets a single CRITICAL system with a severe single-point-of-failure dilute into a "fine" score once there are enough LOW-criticality systems in the same inventory; the weighting keeps that system's risk from disappearing.

All weights live as named constants in RiskRules.cs, each with a one-line rationale.

Every score is auditable

The response does not just give a number — it gives the rules that produced it. Each riskScore ships with a scoreBreakdown of { rule, points, rationale } entries, and

riskScore == min( sum(breakdown.points), 100 )

holds exactly (a test enforces it, so the explanation can never drift from the score). For example:

"singlePointOfFailure": 60,
"scoreBreakdown": {
  "singlePointOfFailure": [
    { "rule": "criticality=CRITICAL", "points": 55, "rationale": "..." },
    { "rule": "type=API",             "points": 5,  "rationale": "..." }
  ]
}

Every response also carries engineVersion and method as provenance. This traceability is the point of choosing deterministic rules over a black box — see METHODOLOGY.md and ARCHITECTURE.md.

Install

dotnet add package CloudSealed.ML.Core

using CloudSealed.ML.Engine.Scoring;
using CloudSealed.ML.Engine.Models;

var response = new ArchitectureAnalyzer().Analyze(request); // PredictArchitectureRequest
Console.WriteLine(response.OverallArchitectureScore);

Use

HTTP service

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. Payloads above ~2MB are rejected.

Response shape:

{
  "predictions": [
    {
      "systemName": "checkout-api",
      "riskScores": { "singlePointOfFailure": 60, "excessiveCoupling": 40, "scalabilityGap": 15 },
      "findings": [
        { "title": "Ponto único de falha: checkout-api", "severity": "HIGH",
          "description": "...", "remediation": "..." }
      ],
      "recommendations": [
        { "title": "Implementar redundância", "description": "...", "effort": "MEDIUM" }
      ]
    }
  ],
  "architectureSummary": "...",
  "overallArchitectureScore": 58
}

CLI

dotnet run --project src/CloudSealed.ML.CLI -- examples/inventory.json
dotnet run --project src/CloudSealed.ML.CLI -- examples/inventory.json --json
dotnet run --project src/CloudSealed.ML.CLI -- examples/inventory.json --html report.html

--html writes a self-contained report (inline CSS, no CDN) alongside whatever other output is requested — open it straight from disk, or attach it to an email.

Runs the same analysis without starting a server, printing either a human-readable summary or the raw JSON response. examples/inventory.json is a ready-to-run sample with a mix of criticality levels and system types.

GitHub Action

Run the audit in CI and get the findings as a pull request comment, without installing anything locally:

- uses: cloudsealed/Predictive-ML-Core@main
  with:
    inventory-json: inventory.json
    fail-on-severity: CRITICAL   # optional: fail the check on CRITICAL findings

Re-runs on the same PR edit the existing comment instead of piling up new ones. See action.yml for all inputs/outputs and .github/workflows/example-usage.yml for a working example (this repository dogfoods its own action against examples/inventory.json on every push).

Alerts

Send the result to Slack (or any generic webhook listener) when a finding reaches a severity threshold, without standing up a dashboard:

dotnet run --project src/CloudSealed.ML.CLI -- examples/inventory.json --webhook-url "$SLACK_WEBHOOK_URL"

A Slack incoming-webhook URL (hooks.slack.com) is auto-detected and rendered as a formatted message; any other URL receives the full JSON response, so it works as-is with Teams, PagerDuty, or a custom listener. Nothing is sent unless a finding is HIGH or CRITICAL. The same behaviour is available in the HTTP API via the optional webhookUrl field on /v1/predict-architecture. A failed webhook is logged and never fails the request.

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 (see ARCHITECTURE.md: "No I/O, no web"). It's the right size when you already have (or can quickly declare) 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.

FAQ

How do I score single-point-of-failure risk for a list of services? Declare each system (name, type, criticality, public exposure, auth method) in a JSON inventory and POST it to /v1/predict-architecture, or run the CLI against the file — see Install and Use.

Why rules instead of a trained model? Because there's no labeled dataset of "this architecture had an incident" to train on, and a model without that data would just wrap heuristics in ML vocabulary — see "Why this is not a trained model".

Can an AI agent call this directly instead of me hitting the API by hand? Yes — see 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.

Development

dotnet restore
dotnet build
dotnet test
dotnet run --project src/CloudSealed.ML.API   # serves on :8092

Tests cover each risk rule in isolation, the criticality-weighted overall score, and the HTTP endpoint (auth, validation, response shape).

License

MIT. See LICENSE.

If the score breakdown helped you argue a redundancy or auth fix, a star helps other teams find it. Bug reports and PRs are welcome — see CONTRIBUTING.md.