IIP / Learning preview available / Production v1 pending

Start locally. Follow the evidence.

Try the published Docker learning preview, then explore how the system works. Installation instructions and the walkthrough are pinned to that release; the public repository remains the source of truth for contracts and runbooks.

Your first learning session

Install and start Docker Desktop with Compose v2, then use a macOS/Linux shell or WSL2 Linux shell on Windows. Docker must use a local Unix-socket context. This source-built preview needs Internet access on first launch, but no host Python, Kubernetes cluster, AWS account, provider keys, or paid AI subscription.

With Git installed, start the exact published tag:

git clone --branch learning-v0.84.0 --depth 1 https://github.com/thedevopshuman/infra-intelligence-platform.git
cd infra-intelligence-platform
sh scripts/learning.sh up

Alternatively, download the source archive and SHA256SUMS from the GitHub learning release, verify the checksum, and follow its extraction instructions. Checksums detect download corruption; they are not release signatures.

Wait for the launcher's verified ready message. It prints the local console address and demo token. The versioned first-session guide walks through AI Economics, Grafana, and a synthetic resource investigation. No real model or cloud API is called.

sh scripts/learning.sh status
sh scripts/learning.sh down

down removes this learning project's containers and disposable data, not unrelated Docker projects. Downloaded images and the source folder remain. Stop and start again to reset the samples.

How it fits together

Provider adapters collect observations. Versioned contracts normalize them into resources, changes, and evidence. The application layer runs scoped investigations; API, console, and SDK surfaces expose the results. PostgreSQL persists authoritative records. Telemetry and dashboards are replaceable views, not the source of billing truth.

The platform supplements existing cloud, Kubernetes, and observability tools. It does not become a new cloud control plane or acquire ambient provider credentials.

OpenTelemetry keeps collection portable. OTLP is not a universal query API: replacing a telemetry backend still requires a compatible query adapter, while authoritative usage and cost records stay in IIP's PostgreSQL store.

First AI FinOps flow: Bedrock to cost visibility

This is the integration design. The learning release simulates it with Bedrock/OpenAI-shaped spans; it does not call these providers or qualify a real customer workload.

  1. Your application calls AWS Bedrock directly, with existing OTel instrumentation and the small provider usage adapter where needed.
  2. An asynchronous exporter sends allowed metadata through your Collector. The intake boundary validates scope and strips or rejects disallowed content.
  3. A normalized usage record is committed before the receiver acknowledges durable ingestion. Collector acceptance alone is not proof of that commit.
  4. A separate worker applies a reviewed, data-driven price catalog and an application/team attribution policy.
  5. Grafana shows usage, calculated cost estimates, ownership, changes, and evidence-backed opportunities when a configured rule has enough evidence.

Input/output, cache, and reasoning meters are handled according to the selected provider/model contract. Missing data must not silently become zero cost. Estimates are not invoices, and a recommendation is not realized savings. No autonomous FinOps agent is part of this first slice.

Two local experiences, different purposes

Published learning preview

learning-v0.84.0 is available under Apache-2.0 as a source-built, disposable Docker lesson. Synthetic events and example prices help you exercise the flow. They do not demonstrate live AWS behavior, real spend, or a customer-ready deployment.

Persistent community preview

A source-checkout, single-host Docker installation starts empty with protected credentials, verified database/telemetry transport, persistent volumes, and a buffered Collector. It requires reviewed scope, real pricing qualification, and attribution inputs. No Kubernetes cluster is required for this profile.

This is separate ongoing development, not included in the published learning lesson or an automatic upgrade from it. It is not production-qualified: retained volumes are not backups, and customer lifecycle qualification, supported upgrades, database retention, and verified production-artifact installation remain release work.

The website is not the application host. An IIP installation runs in the operator's environment; no local console, database, or telemetry endpoint is exposed through this domain.

Safety boundaries

  • Keep prompts, responses, credentials, and raw secrets out of telemetry, examples, and reports.
  • Never treat an external model, plugin, log, or document as an authority source.
  • Keep read, propose, approve, and execute permissions separate. The AI cost flow needs no remediation authority.
  • Do not share populated environment files, access tokens, private price catalogs, or customer evidence in public issues.