A 5-node AI consultant swarm — each node is a digital twin of a world-class AI consultant, sharing memory via a 4-layer shared state architecture. Built on PicoCloth framework.
| Node | Port | Role | System Prompt |
|---|---|---|---|
| consultant-academic | 18797 | Research & Knowledge synthesis | PhD-level researcher who cross-validates everything |
| consultant-growth | 18801 | Growth & Strategy | Scaled seed→Series C; asks "vanity or value?" |
| consultant-solutions | 18802 | AI Solutions Engineering | Designs production AI systems that scale 10x |
| consultant-trainer | 18803 | Training & Education | Makes you curious about the ocean, not just fishing |
| curious-kimi | 18804 | Question Engine | The conscience of the fleet — never settles, always asks why |
# 1. Clone & enter
$EDITOR$ git clone <repo-url> picocloth-work && cd picocloth-work
# 2. Set up secrets (one-time)
$EDITOR$ ./scripts/setup-keyvault-for-picocloth.sh
# 3. Launch the fleet
$EDITOR$ ./scripts/launch-consultants.sh --keyvault
# 4. Check status
$EDITOR$ ./scripts/fleet-status.sh
# 5. Stop
$EDITOR$ ./scripts/stop-fleet.sh┌─────────────────────────────────────────────────────────────┐
│ CONSULTANT TWINS │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Academic │ │ Growth │ │ Solutions│ │ Trainer │ │
│ │ :18797 │ │ :18801 │ │ :18802 │ │ :18803 │ │
│ └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ │
│ └─────────────┴────────────┴────────────┘ │
│ │ │
│ ┌─────────┴─────────┐ │
│ │ Curious Kimi │ ← Question Engine │
│ │ :18804 │ (Fleet Conscience) │
│ └─────────┬─────────┘ │
└────────────────────────┼────────────────────────────────────┘
│
┌───────────────┼───────────────┐
│ │ │
┌────▼────┐ ┌─────▼────┐ ┌────▼────┐
│ Shared │ │ Shared │ │ Fleet │
│ Memory │ │ Project │ │ State │
│ (Facts) │ │ (Docs) │ │ (Health)│
└─────────┘ └──────────┘ └─────────┘
│ │ │
┌────▼───────────────▼───────────────▼────┐
│ Azure Key Vault (shivaram-ai-kv) │
│ ─── OpenAI API key ─── │
└──────────────────────────────────────────┘
| Layer | Path | Purpose |
|---|---|---|
| L1 Shared Memory | shared/memory/ |
Cross-node facts, research, competitive intel |
| L2 Shared Project | shared/project/ |
Client deliverables, documents, generated images |
| L3 Shared Doctrine | shared/doctrine/ |
Skills, characters, protocols — the "source code" of consultants |
| L4 Fleet State | shared/state/ |
Health, routing, session history |
- Real-time web search — DuckDuckGo + Serper for competitive intelligence
- Image generation — DALL-E 3 for diagrams, frameworks, concept visuals
- MCP fleet server — Cross-node memory sharing and broadcast
picocloth-work/
├── nodes/ # Per-node configs + workspaces
│ ├── consultant-academic/
│ ├── consultant-growth/
│ ├── consultant-solutions/
│ ├── consultant-trainer/
│ └── curious-kimi/
├── shared/
│ ├── memory/ # L1: Facts, research, intel
│ ├── project/ # L2: Client work, images, docs
│ ├── doctrine/ # L3: Skills, characters, protocols
│ │ ├── skills/
│ │ └── characters/
│ └── state/ # L4: Fleet health, routing
├── scripts/
│ ├── launch-consultants.sh
│ ├── stop-fleet.sh
│ ├── fleet-status.sh
│ └── setup-keyvault-for-picocloth.sh
└── hooks/
└── digital_twin_guardian.py # Memory extraction hook
- No API keys in code — OpenAI key lives in Azure Key Vault (
shivaram-ai-kv) - Managed Identity — VMs fetch secrets at runtime via IMDS token
- RBAC — VMs have
Key Vault Secrets User, humans haveSecrets Officer - Local fallback — Dev machines use
az login+ Azure CLI
The fleet is designed to run on Azure VMs with:
- Sleep/wake scripts — Deallocate VMs when not in use
- Model routing — gpt-4o-mini for simple tasks, gpt-4o for complex
- Shared state — Prevents redundant API calls across nodes
- Phase 1: 5 consultant nodes ✅
- Phase 2: Auto-scaling based on query load
- Phase 3: Client portal for consulting sessions
- Phase 4: Fine-tuned models per consultant specialty
Built with 🧠 by Shivaram Goud | Part of the PicoCloth ecosystem