AI代理编排平台

ai-platform

AI Platform POC — agent-api, rag-service, mcp-server, n8n orchestration on AWS EKS

Platform SSOT

• platform-config.yaml — Bedrock model IDs / embedding model (LLM SSOT).
• packages/chat_helper/pricing.py — chat caps. MAX_OUTPUT_TOKENS defaults to 400 and reads env MAX_OUTPUT_TOKENS. fit_input trims input to MAX_INPUT_TOKENS (8000). Infrastructure sets the Lambda env to 400 and reserved concurrency to 2. The account-wide Bedrock alarm (Sum > 30 Invocations per 300 seconds) lives in LevKesha/infrastructure and counts every Bedrock call, including chat, RAG, and LiteLLM.
• platform-infra.dev.json — infra IDs snapshot (ACM, subnets, IRSA, ECR). Owner: LevKesha/infrastructure terraform outputs; refresh via ../infrastructure/scripts/export-platform-outputs.py.
• packages/platform_common/ — shared JSON logging, AWS text summaries, Bedrock Claude invoke. Vendor into service repos: python scripts/sync-platform-common.py (edit SSOT only; do not hand-edit copies).
• Render k8s from snapshot: python scripts/render-from-infra.py
• Sync sibling helm/agent-spec IRSA+ECR: python scripts/sync-sibling-irsa.py
• Drift checks: python scripts/check-platform-config.py and python scripts/check-infra-consumers.py
• Path ownership: python3 scripts/check-ownership.py (ownership.json)
• Security review for auth, secrets, identity, and the public edge: python3 scripts/check-security-review.py

Theme 4 status

• n8n-config/ contains the orchestrator workflow JSON + usage docs.
• n8n/ contains a self-hosted EKS baseline for n8n + Postgres.