Production-grade, microservices, secure-by-default, GENAI-ready — not a scaffold. Stack: FastAPI (Python 3.11) + Next.js 14 + Postgres 16/Timescale + Redis + NATS + Qdrant + Tailwind + Docker + GKE Autopilot + Terraform + GitHub Actions OIDC Guards: Raw frames NEVER leave plant (type-enforced) · Dual reasoning (Nemotron-9B on-prem + Gemini ER2 cloud) · DPDP 2023 asia-south1 · OWASP ASVS 4.0 + SOC2
Skopaq AI Private Limited · Product family sibling to VELA · Working codename TANTU (Sanskrit: thread) — uncleared, externals use Skopaq AI until IP-India classes 9,42,7 + MCA pass.
TANTU makes every machine on a mixed-vendor floor speak one language, watch itself, and answer in the operator's language — for a monthly fee a mid-market Indian plant (200–2,000 workers, PLI-era) can approve without a board meeting.
Why not a platform to out-Siemens Siemens: Siemens Industrial Copilot (120K+ engineers, BMW) and Rockwell FactoryTalk Copilot (Nemotron-9B on HMI, Ford) already own the horizontal copilot. They structurally cannot do: (1) cross-vendor orchestration, (2) Indian mid-market price, (3) vernacular voice-first hi/ta/te/kn, (4) legacy retrofit (camera on 1990s gauges). TANTU lives there.
Pricing: low-tens K Rs /machine-cluster/mo (modeled, validate in 3 pilots) — opex, reversible, 90-day pilot; hardware Rs 40–80K amortized.
[Factory Floor: Fanuc/ABB/KUKA/Yaskawa/UR + Siemens/Omron PLCs + analog gauges]
↓
CONNECT adapter-fabric:8001 — OPC-UA (asyncua) · Modbus TCP/RTU (pymodbus) · MQTT (paho) · MTConnect (xml) · Eth/IP (pycomm3) · camera-as-adapter (OpenCV Hough)
→ normalized NormalizedReading/DefectEvent (NO image field → frames never leave) — tag-map compounding lib
↓ NATS
PERCEIVE edge-perception:8002 — tiered Pi5+Hailo-8L / Orin Nano / Thor, gauge CV (CLAHE+Hough) + vibration FFT (numpy/scipy) + thermal/CT, <40ms, store-and-forward Redis, OTA
↓ derived events only
REASON reasoning-copilot:8003 — dual: Gemini ER2 (google-genai SDK) on cloud + Nemotron-9B (vLLM/Ollama) on-prem (air-gapped flag) → Qdrant RAG (sentence-transformers) + prompt registry + vernacular TTS/STT
↓ CorrelationReport
ORCHESTRATOR 8004 — window policy 2 faults or conf≥0.97 → escalate, NATS, TimescaleDB, idempotency
↓
GATEWAY 8000 — JWT RS256 + RBAC/ABAC (plant_id) + rate-limit + OPA + audit, proxies to 8001-8004
FRONTEND 3000 — Next.js 14, 3-role (Operator voice-first, Maintenance SSE grid, Plant-Head opex), 5-lang, Recharts
Diagram: Factory → [VPC private service networking] → GKE Autopilot (Gateway API L7 + Argo Rollouts canary 20→50→100) → Postgres private HA + Redis private + Qdrant Helm → Artifact Registry (GHCR+GAR).
World-class patterns: Hexagonal + Clean + DDD + Event-Driven + CQRS-lite + Saga (backend) · 12-Factor + OpenTelemetry + Prometheus/Grafana/Loki · RAG + Grounded + Hallucination guard · Offline-First + Tiered.
tantu-platform/
services/adapter-fabric/ FastAPI :8001, 6 adapters, domain/events.py, tests, Dockerfile
services/edge-perception/ FastAPI :8002, gauge+vibration+thermal, tiered, tests
services/reasoning-copilot/ FastAPI :8003, Gemini+vLLM+Qdrant+RAG+vernacular, tests
services/orchestrator/ FastAPI :8004, NATS+policy, Timescale, tests
services/api-gateway/ FastAPI :8000, proxy+auth+rate-limit, tests
frontend/ Next.js 14 :3000, app/{operator,maintenance,plant-head}, components/ui, lib/i18n, hooks/useSpeech+useSSE
simulation/ Digital twin stub + ScenarioRunner
terraform/ GKE Autopilot + VPC + Cloud SQL 16/Timescale + Redis + Registry + Qdrant + IAM + Secrets + Cloud Build
envs/{dev,prod}.tfvars asia-south1, cost-optimized (Spot dev, 3× HA prod)
infra/k8s/ Deployment+HPA+Gateway+HTTPRoute+Rollout (Argo) + ExternalSecrets
.github/workflows/{ci,cd,release,infra}.yml + .github/release.config.js
scripts/{version.sh,gcp-beta-apply.sh,seed.py}
docs/{ADR-001,THREAT_MODEL,DPDP_MEMO,VERSIONING}.md
docker-compose.microservices.yml 5 services + 4 infra (postgres/redis/nats/qdrant) + frontend
.version + services/*/VERSION single-source SemVer
git clone https://github.com/Skopaq-AI/tantu-platform.git && cd tantu-platform
cp .env.example .env # optional: add GEMINI_API_KEY
# microservices (recommended):
docker compose -f docker-compose.microservices.yml up --build -d
# verify:
curl http://localhost:8001/health # adapter-fabric
curl http://localhost:8002/health # edge
curl http://localhost:8003/health # reasoning
curl http://localhost:8004/health # orchestrator
curl http://localhost:8000/health # gateway (aggregated)
open http://localhost:3000 # frontend — 3 tabs
open http://localhost:8000/docs # gateway OpenAPI
# demo: 3 protocols → 1 schema → 1 vernacular answer
python demo.py
curl -X POST http://localhost:8000/api/v1/poll -H "Content-Type: application/json" -d '{"protocol":"camera","station_id":"line2-cluster1-gauge3","tag":"gauge-3"}' | jq
curl -X POST http://localhost:8000/api/v1/ask -H "Content-Type: application/json" -d '{"question":"why did line 2 slow down?","plant_id":"plant-demo-01","lang":"ta"}' | jq
make seed # plant-demo-01 + tag-map + 20 gaugesAccounts (seeded, JWT): operator/operator123 (ta), maintenance/maint123 (en), plant_admin/admin123 (en) — all plant_id=plant-demo-01.
In GCP console (project tantu-beta-20260821-01 (dedicated beta project), region asia-south1):
# variables
export PROJECT_ID=tantu-beta-xxxxx
export REGION=asia-south1
export GH_REPO=Skopaq-AI/tantu-platform
# enable APIs
gcloud services enable iamcredentials.googleapis.com sts.googleapis.com --project $PROJECT_ID
# Workload Identity Pool + Provider (GitHub OIDC) — see terraform/modules/iam (idempotent)
gcloud iam workload-identity-pools create github --project $PROJECT_ID --location global --display-name "GitHub" 2>&1 | head
gcloud iam workload-identity-pools providers create-oidc github-provider --project $PROJECT_ID --location global \
--workload-identity-pool github --issuer-uri https://token.actions.githubusercontent.com \
--attribute-mapping "google.subject=assertion.sub,attribute.repository=assertion.repository,attribute.ref=assertion.ref" \
--attribute-condition "assertion.repository==\"$GH_REPO\"" 2>&1 | head
# Service account for Actions
gcloud iam service-accounts create gha-deployer --project $PROJECT_ID --display-name "GitHub Actions Deployer"
gcloud projects add-iam-policy-binding $PROJECT_ID --member "serviceAccount:gha-deployer@$PROJECT_ID.iam.gserviceaccount.com" --role "roles/artifactregistry.writer"
gcloud projects add-iam-policy-binding $PROJECT_ID --member "serviceAccount:gha-deployer@$PROJECT_ID.iam.gserviceaccount.com" --role "roles/container.admin"
gcloud projects add-iam-policy-binding $PROJECT_ID --member "serviceAccount:gha-deployer@$PROJECT_ID.iam.gserviceaccount.com" --role "roles/cloudsql.editor"
gcloud iam service-accounts add-iam-policy-binding gha-deployer@$PROJECT_ID.iam.gserviceaccount.com \
--project $PROJECT_ID --role "roles/iam.workloadIdentityUser" \
--member "principalSet://iam.googleapis.com/projects/$(gcloud projects describe $PROJECT_ID --format="value(projectNumber)")/locations/global/workloadIdentityPools/github/attribute.repository/$GH_REPO"Then in GitHub: Settings → Secrets and variables → Actions:
- Secrets:
GCP_WIP=projects/123456/locations/global/workloadIdentityPools/github/providers/github-provider,GCP_SA=gha-deployer@<PROJECT_ID>.iam.gserviceaccount.com,GEMINI_API_KEY(optional) - Variables:
GCP_PROJECT_ID=tantu-beta-xxxxx,GCP_REGION=asia-south1,GCP_GAR_REPOSITORY=tantu
Terraform remote state bucket (once): gsutil mb -l $REGION gs://$PROJECT_ID-tfstate
- CI (
.github/workflows/ci.yml, onpushtomain+ PR): matrixruff+mypy+pytestper service +eslint/tsc/vitestfrontend +pip-audit/npm audit/gitleaks+terraform fmt/validate. Caching viaactions/cache. Branch protection: requireCIcheck onmain. - CD (
.github/workflows/cd.yml, onpush:mainafterCIsuccess +workflow_dispatch+ tagsv*):buildx6 services (adapter-fabric:8001…frontend:3000) → GHCRghcr.io/Skopaq-AI/tantu-platform/<svc>+ GCP Artifact Registryasia-south1-docker.pkg.dev/<PROJECT_ID>/tantu/<svc>with tags0.1.0-<sha>,latest(onmain),sha,semver. OIDC viagoogle-github-actions/auth@v2, cosign keyless sign + Syft SBOM, provenance false, cachegha. - Release (
.github/workflows/release.yml, onpush:main): Conventional Commits →semantic-release(feat→minor,fix→patch,BREAKING CHANGE→major) → bumps.version+services/*/VERSION+pyproject.toml+frontend/package.jsonviascripts/version.sh, writesCHANGELOG.md, createsvX.Y.Ztag + GitHub Release withinfra/k8s/*.yamlassets. - Infra (
.github/workflows/infra.yml, onpush:mainpathterraform/**orworkflow_dispatchwithenv):terraform init(GCSgs://$PROJECT_ID-tfstateprefixenv/dev) →plan→apply(auto-approve onmain), thengcloud container clusters get-credentials+helm upgrade --install(values per env) +kubectl applyGateway API. Uses same OIDC.
Why setup from GitHub Actions failed before: GCP_WIP/GCP_SA/GCP_PROJECT_ID vars were empty and the sandbox proxy blocked oauth2.googleapis.com + registry.terraform.io. Fix is the 1-time OIDC pool above + set the 3 vars. Once set, CD pushes to GAR and Infra can terraform apply from Actions (no local gcloud needed). Logs at Actions → CD / Infra.
Version locally: ./scripts/version.sh get / bump patch / check — full guide docs/VERSIONING.md.
Local (outside sandbox, where gcloud works):
./scripts/gcp-beta-apply.sh tantu-beta-$(date +%s) asia-south1 dev
# does: gcloud projects create → services enable → gsutil mb tfstate → terraform init/plan/apply (envs/dev.tfvars) → gke auth → helm upgrade + gatewayVia GitHub Actions (after OIDC vars set): Actions → infra → Run workflow → env: dev → Run (or push to terraform/** on main).
Cost: dev ~$80/mo (Autopilot + 1× db-custom-2-7680 + 1GB Redis + Spot), prod (values-prod.yaml) 3× HA ~$450/mo.
make test # pytest 5 services (30/30) + vitest frontend
make lint # ruff + mypy + eslint + tsc
make audit # pip-audit + npm audit + gitleaks- Secrets via
Secret Manager/infra/vault(never in env),gitleaksin CI - JWT RS256 (HS256 dev fallback) + RBAC (operator/maintenance/plant_admin) + ABAC (
plant_id) — seeservices/*/src/*/infra/security.py - Rate limit (Redis) + CORS + Helmet + OPA-style policy
- TLS 1.3, at-rest via Cloud SQL TDE, field-level PII, DPDP
data_residency=IN, 90d hot/1y cold, audit log (Timescale), raw frames never leave (type-enforcedNormalizedReadinghas noimagefield)
reasoning-copilot:8003routes byair_gapped:google.genai.Client(api_key=GEMINI_API_KEY).models.generate_content(model="gemini-robotics-er-2", ...)vshttpxtoVLLM_URL(Nemotron-9B). Prompt registryv1, RAG Qdrant (sentence-transformers), grounded + citations, hallucination guard.
Codename TANTU is uncleared — externals use Skopaq AI until IP-India + MCA pass. All pricing modeled pending pilot.