diff --git a/README.md b/README.md index dc443a0..2d12d42 100644 --- a/README.md +++ b/README.md @@ -22,7 +22,7 @@ A production-grade cloud-native infrastructure laboratory demonstrating GitOps, This repository is the single source of truth for a multi-application cloud platform running on GKE. Every deployment, database, and network policy flows through Git via Flux CD. What lives here: -1. **AgentForge** — Private multi-tenant AI agent workspace with dual-tier vLLM inference (L4 dispatcher + A100 deep thinker) and isolated CNPG databases per tenant. +1. **AgentForge** — Private multi-tenant AI agent workspace with dual-tier vLLM inference (L4 dispatcher + RTX 6000 deep thinker) and isolated CNPG databases per tenant. 2. **Multi-Profile AI Agent Team** — Six specialist AI profiles (backend-dev, frontend-dev, researcher, outreach, quant, sec-ops) orchestrated through a shared Kanban board with automated audit-to-fix pipelines. 3. **Waitlist API** — FastAPI landing page backend with idempotent signups, async PostgreSQL, and Telegram fire-and-forget notifications. 4. **Autonomous News Quant Pipeline** — 371 global feed scraper with DeepSeek-R1 analysis generating actionable futures trading signals. @@ -40,8 +40,8 @@ This repository is the single source of truth for a multi-application cloud plat │ │ GKE Cluster (devops-lab-cluster) │ │ │ │ │ │ │ │ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │ │ -│ │ │ Standard │ │ L4 GPU Pool │ │ A100 GPU Pool│ │ │ -│ │ │ Node Pool │ │ (SPOT L4) │ │ (SPOT A100) │ │ │ +│ │ │ Standard │ │ L4 GPU Pool │ │ RTX 6000 GPU │ │ │ +│ │ │ Node Pool │ │ (SPOT L4) │ │ (SPOT RTX6K) │ │ │ │ │ │ e2-std-2 │ │ 1 node (24/7)│ │ 0-1 nodes │ │ │ │ │ └──────────────┘ └──────────────┘ └──────────────┘ │ │ │ │ │ │ @@ -56,10 +56,10 @@ This repository is the single source of truth for a multi-application cloud plat │ │ │ │ │ │ ┌───────────────────────────────┐ ┌───────────────────────────────┐ │ │ │ │ │ customer1 namespace │ │ agent-forge namespace │ │ │ -│ │ │ - AgentForge (PAaaS) │ │ - Tenant-specific OpenClaw │ │ │ +│ │ │ - AgentForge (PAaaS) │ │ - Tenant-specific Hermes Agent │ │ │ │ │ │ - Dual-tier vLLM │ │ - Isolated CNPG databases │ │ │ │ │ │ L4 Dispatcher (24/7) │ │ - Qwen 3.6 27B Abliterated │ │ │ -│ │ │ A100 Deep Thinker (KEDA) │ │ - KEDA scale-to-zero │ │ │ +│ │ │ RTX 6000 Deep Thinker (KEDA) │ │ - KEDA scale-to-zero │ │ │ │ │ │ - Waitlist API (FastAPI) │ │ │ │ │ │ │ │ - News Bot Pipeline │ │ ┌───────────────────────────┐ │ │ │ │ │ │ - Landing Page │ │ │ sec-ops audit agent │ │ │ @@ -128,10 +128,10 @@ This repository is the single source of truth for a multi-application cloud plat | Tool | Version | Purpose | |------|---------|---------| | **vLLM** | v0.9.1 | High-throughput LLM inference server | -| **Qwen 3.6 27B Abliterated** | Latest | Uncensored reasoning model (A100 deep thinker tier) | +| **Qwen 3.6 27B Abliterated** | Latest | Uncensored reasoning model (RTX 6000 deep thinker tier) | | **Qwen 2.5 Coder 7B Abliterated** | Latest | Fast tool-calling dispatcher (L4 24/7 tier) | | **NVIDIA L4 GPU** | - | 24/7 GPU for fast triage and dispatch | -| **NVIDIA A100 80GB** | - | SPOT GPU for deep reasoning and multi-file context | +| **NVIDIA RTX 6000 Pro** | - | SPOT GPU for deep reasoning and multi-file context | ### Development Environment @@ -160,7 +160,7 @@ gcloud-lab/ │ ├── gke.tf # GKE cluster definition │ ├── vpc.tf # VPC and subnet configuration │ ├── nodepool.tf # Standard node pool -│ ├── nodepool-gpu.tf # GPU node pools (L4 + A100 SPOT) +│ ├── nodepool-gpu.tf # GPU node pools (L4 + RTX 6000 SPOT) │ ├── flux.tf # Flux GitOps bootstrap │ ├── helm.tf # Helm chart deployments (Cilium) │ └── variables.tf # Input variables @@ -213,8 +213,8 @@ gcloud-lab/ │ │ │ └── scrapy-urls-configmap.yaml │ │ ├── agent-forge/ │ │ │ ├── namespace.yaml -│ │ │ ├── vllm-deep-thinker.yaml # A100 deployment with KEDA -│ │ │ ├── openclaw-tenant.yaml # Per-tenant OpenClaw instance +│ │ │ ├── vllm-deep-thinker.yaml # RTX 6000 deployment with KEDA +│ │ │ ├── hermes-tenant.yaml # Per-tenant Hermes agent instance │ │ │ └── pg-cluster-agentforge.yaml │ │ ├── kanban/ │ │ │ ├── namespace.yaml @@ -259,7 +259,7 @@ gcloud-lab/ |------|-------------|---------|---------| | Standard | e2-standard-2 | 1-16 nodes | General workloads, N8N, web servers | | GPU L4 (SPOT) | g2-standard-8 + L4 | 0-5 nodes | vLLM dispatcher, 24/7 fast inference | -| GPU A100 (SPOT) | a2-highgpu-1g + A100 80GB | 0-1 nodes | Deep thinker tier, multi-file reasoning | +| GPU RTX 6000 (SPOT) | g6-standard-4 + RTX 6000 Pro | 0-1 nodes | Deep thinker tier, multi-file reasoning | ### Networking @@ -279,7 +279,7 @@ Multiple isolated PostgreSQL clusters, each with dedicated databases per applica |---------|-----------|-----------|--------| | `customer1-pgdb` | customer1 | `n8n`, `news_app`, `waitlist` | GCS, 7-day retention | | `hermes-pgdb` | kanban | `hermes`, `memory_store` | GCS, 7-day retention | -| `openclaw-pgdb` | agent-forge | Per-tenant isolated DBs | GCS, 7-day retention | +| `hermes-tenant-pgdb` | agent-forge | Per-tenant isolated DBs | GCS, 7-day retention | | `siriusdevops-pgdb` | customer1 | `waitlist_prod` | GCS, 30-day retention | ### GitOps Flow @@ -310,9 +310,9 @@ GitHub Repository (ghcr.io/sirius0xdev) A premium, uncensored, privacy-first AI agent hosting platform with dual-tier cognitive architecture: - **Tier 1 (Dispatcher):** L4 GPU SPOT instance running 24/7. Hosts `Qwen2.5-Coder-7B-Instruct-heretic` via vLLM `v0.9.1` for lightning-fast, cheap triage and tool calling. -- **Tier 2 (Deep Thinker):** A100 80GB SPOT instance scaling from 0-1 via KEDA. Hosts `Qwen3.6-27B-heretic` with chunked prefill and FP8 KV cache for massive multi-file context and reasoning without OOMing. -- **Multi-Tenant Isolation:** Each tenant gets an isolated OpenClaw deployment with its own CNPG PostgreSQL database. No cross-tenant data leakage. -- **Landing Page:** Dockerized marketing site at siriusdevops.com, built via CI/CD from GitHub Actions and deployed to the staging kustomization overlay. +- **Tier 2 (Deep Thinker):** RTX 6000 Pro Spot instance scaling from 0-1 via KEDA. Hosts `Qwen3.5-27B-heretic` with `--enable-chunked-prefill` and `--kv-cache-dtype=fp8` for massive multi-file context and reasoning without OOMing or stalling concurrent users. +- **Frontend:** Isolated Hermes agent profiles per tenant, connected to Telegram/Discord via outbound polling (no public ingress required). +- **Landing Page:** Dockerized marketing site built via CI/CD from `hermes-projects` and deployed to the `staging` kustomization overlay. - **Container Registry:** All images pushed to `ghcr.io/sirius0xdev`. ### 2. Multi-Profile AI Agent Team @@ -463,7 +463,7 @@ or create a `monitoring-grafana-admin` Secret instead. ### Database Security - Managed roles with secret-based passwords per application -- Separate PostgreSQL clusters per domain (hermes-pgdb, openclaw-pgdb, siriusdevops-pgdb) +- Separate PostgreSQL clusters per domain (hermes-pgdb, hermes-tenant-pgdb, siriusdevops-pgdb) - GCS backups with configurable retention policies - HA cluster with automatic failover @@ -477,8 +477,8 @@ or create a `monitoring-grafana-admin` Secret instead. ## Cost Optimization -- **SPOT GPU Instances**: 60-90% savings on L4 and A100 workloads -- **KEDA Scale-to-Zero**: A100 deep thinker pool scales to 0 when no requests are queued +- **SPOT GPU Instances**: 60-90% savings on L4 and RTX 6000 workloads +- **KEDA Scale-to-Zero**: RTX 6000 deep thinker pool scales to 0 when no requests are queued - **Resource Limits**: CPU and memory caps on every container prevent runaway costs - **Scheduled Workloads**: CronJobs only run when needed — no idle inference pods - **Tailscale for Internal Access**: No need for expensive internal load balancers or Cloud NAT for monitoring