Screenshots

All shots are from live clusters — mostly a local Kind cluster brought up with ./scripts/setup.sh + ./scripts/bootstrap-ai.sh, plus a few from the AWS EKS path. No mock-ups.

Some tabs show live data from third-party integrations (SonarCloud, Datadog, PagerDuty, …) that will be empty on a fresh install until you add your own credentials — see Optional third-party integrations.

The portal

The Platform Dashboard is the landing page: catalog counts, platform-wide DORA, and every service at a glance.

Software Catalog	Teams
Every service, API, MCP server and test suite, owned and tagged	8 teams, each with its own namespace, SecretStore and Grafana folder

Each entity page carries the platform's own tabs — TechDocs, Kubernetes, DORA, Scorecard, Security, Datadog, Trivy, SLOs:

TechDocs on the entity	Scorecard on the entity

Golden path — scaffold → repo → deploy

59 templates in the Scaffolder, filtered by category, tag or owner:

Scaffolder task running	The repo it produced
Generate → push to GitHub → register in catalog → run the first job	CI workflow, Dockerfile, catalog-info.yaml, TechDocs — all wired

Templates are wizards, not a wall of fields — and the last step opens the GitOps PR that puts the new service under ArgoCD:

Template wizard (LLM App)	The GitOps PR it opened
Model, effort level, trace sampling — chosen up front, wired into the skeleton	An ApplicationSet auto-discovers the service into services-dev on merge

Self-service infrastructure is the same flow — a Crossplane Claim committed to Git instead of a Terraform PR:

ArgoCD app-of-apps	Argo Rollouts canary

Shift-left quality

Bronze / Silver / Gold tiers across every service, with the cheapest unfilled check called out as the next action:

SLOs and error budgets	QA platform metrics
Sloth multi-window burn-rate, live from Prometheus	E2E pass rate, k6 p95 latency and error rate per run

Those SLOs aren't hand-written YAML — a template generates the Sloth definitions and burn-rate alerts and opens the PR:

AI/ML platform and agents

The AI Assistant answers in plans, not prose — it maps your intent onto the actual templates on the platform and asks for exactly the inputs they need:

Ask it anything	It plans the scaffold

…and then it actually runs the scaffolder — repo, deploy target, task ID and the suggested next steps come back in the same chat:

KAgent agents	MCP servers and model configs

The MLflow page gives experiment tracking and the model registry the same in-portal treatment — experiments, recent runs and registered models, read live from the MLflow API, without leaving the catalog:

Agent Approvals (HiTL gate)	MLflow's own UI
Every mutating agent action waits for a human — or an auto-approve policy	One click away at mlflow.idp.local, for the deep-dive views

Semantic search over templates, components and TechDocs (Voyage AI + pgvector):

Prometheus tells you that an agent ran. Langfuse tells you what it cost — prompt and completion, tokens, latency and spend per run, without leaving the portal:

AI Observability in Backstage	The full Langfuse UI
Cost and token usage per model, and recent agent runs with latency	Cost by model and environment, per-user spend, and trace drill-down

Observability, DORA and FinOps

DORA metrics	FinOps cost overview
Four keys platform-wide and per service, with performance bands	OpenCost spend by namespace, team or container

Cost Calculator	Grafana — IDP services
Estimate a service's monthly cost before scaffolding it	Request rate, CPU/memory and restarts, filtered by catalog entity

Incidents are records, not Slack threads — auto-filed from Alertmanager, severity-filtered, and feeding MTTR back into DORA:

More screens — Tech Radar, onboarding, Learning Center, API explorer, Copilot metrics, admin, activity feed, search, support

Tech Radar (92 entries)	API Explorer

OpenAPI spec on the entity	Onboarding

Learning Center	Copilot metrics

Admin	Activity feed

Search	Support