hetu (हेतु, Sanskrit) — cause, reason, motive. The thing this engine tries to find when something goes wrong in your cluster.

A Kubernetes cluster intelligence engine: deterministic health scoring, four-level drill-down, Sentry-style error grouping, anomaly detection, and LLM-backed root-cause analysis. Built in Go (collector + analyzer) and Next.js 16 (dashboard).

Ops teams drown in green-dot telemetry. Existing tooling shows you raw counts but not impact, hides scoring inside black boxes, and stops drill-down at "open Grafana". Errors stay open for weeks; memory leaks.

hetu answers a different question: "why is this pod hurting my score, and what should I do about it?" Every number on the dashboard is auditable to a rule and drills down four levels deep — from a top-level score, into the rule that deducted it, into the resources causing the deduction, into the per-resource impact tab.

Overview — health scores + AI insights	Errors — Sentry-style grouping

Errors detail — sparkline + spike + severity chip	Workloads browser

Incidents & RCA	Anomalies (z-score)	Recommendations

Three Go services + a Next.js dashboard, glued by HTTP/SSE.

┌──────────────┐    ┌────────────────┐    ┌──────────────┐    ┌──────────────┐
│ Kubernetes   │ →  │   collector    │ →  │   analyzer   │ →  │  dashboard   │
│  + Prom      │    │  Go · ring 10K │    │  Go · scoring│    │  Next.js 16  │
└──────────────┘    │  qps 50 burst  │    │  + RCA + LLM │    │  + SSE proxy │
                    │  100 · resync  │    │  + eviction  │    └──────────────┘
                    │  5m · scrape   │    │  loop        │
                    │  30s           │    └──────────────┘
                    └────────────────┘

Full diagram & component breakdown: see the impress.js deck at docs/presentation/index.html (open in any modern browser).

Every number below is in the source. Italic = file & line.

Requires Go 1.25, Node 22, a reachable Kubernetes context, and (optionally) an Ollama / OpenAI / Anthropic endpoint for LLM features.

# Interactive runner — prompts for every config, validates, then starts
# collector + analyzer + dashboard.
scripts/run-local.sh

# Or non-interactive (CI shape):
ENVIRONMENT=dev scripts/run-local.sh start --yes

# Pre-flight without starting anything
scripts/run-local.sh doctor

# Lint env file only
scripts/run-local.sh lint

# Stop, status, restart, logs (per-service tail)
scripts/run-local.sh {stop|status|restart|logs}

Dashboard at http://localhost:3003, analyzer at http://localhost:18081, collector at http://localhost:18080/healthz.

Full reference: docs/script_usage.md.

helm upgrade --install hetu deploy/helm/cluster-intel \
  --namespace cluster-intel --create-namespace \
  -f values-deploy.yaml

In-cluster guide: deploy/helm/cluster-intel/README.md.

• Health scoring math — docs/SCORING_SYSTEM.md
• Confidence honesty — docs/CONFIDENCE_SCORES.md (why some confidence numbers don't mean what you think)
• Architecture deep-dive — docs/ARCHITECTURE.md
• API contracts — docs/API_CONTRACTS.md
• Errors plan (in progress) — docs/ERRORS_PLAN.md
• Demo walkthrough — docs/DEMO_WALKTHROUGH.md
• Roadmap — docs/ROADMAP.md

This is a working MVP — 196 tests, 0 failures, deterministic scoring, in-memory storage with TTL eviction. It is not yet wired to a persistence layer; restart loses incidents/errors/anomalies. Work in flight (see docs/ROADMAP.md and docs/ERRORS_PLAN.md):

• Postgres persistence + score trendlines
• Errors: rate buckets ✅ · severity sort ✅ · pagination ✅ · cross-service faultKey · embedding-based dedup · typed LLM Analysis
• CI workflow (GitHub Actions)
• Multi-cluster aggregation
• Pluggable scoring rules

Bug reports and PRs welcome. Before opening a PR:

go test ./...                                    # Go suite
cd src/dashboard && npm test && npx playwright test  # JS / E2E
scripts/run-local.sh doctor                      # local pre-flight

MIT