Starlight Intelligence System (SIS)
Strategic AI orchestration layer with multi-model routing, agent coordination, and purpose-driven decision intelligence
The Problem
Multi-agent systems lack strategic coordination. Individual agents optimize locally but miss global context. Model routing is static (always Claude or always GPT), wasting cost on simple tasks and quality on complex ones. There's no learning layer that captures what worked and applies it to future decisions.
The Solution
Deploy a strategic orchestration layer with intent classification (16 categories), dynamic model routing based on task complexity and budget, a reasoning bank that learns from past trajectories, and an agent coordination protocol that enables agents to hand off context without losing state.
Overview
The Starlight Intelligence System is a meta-orchestration layer that sits above individual agents and coordinates them toward strategic outcomes. It routes requests to optimal models (Claude, Gemini, GPT, local LLMs), maintains a shared reasoning bank for pattern learning, and provides a unified intelligence API for all downstream applications. Think of it as the brain's prefrontal cortex — it doesn't do the work, it decides what work needs doing and who does it best.
Architecture
Components
SIS Gateway
gatewayUnified API endpoint for all intelligence requests. Classifies intent, determines complexity, and routes to optimal processing path.
Service: Vercel Edge Functions
Intent Classifier
ai-service16-category intent classification. Maps requests to domains: code, research, creative, strategy, health, communication, analysis, etc.
Service: Claude Haiku (fast)
Model Router
computeSelects optimal model based on task complexity, budget tier, and historical performance. Routes between Claude Opus, Sonnet, Haiku, Gemini, GPT, and local LLMs.
Service: Custom routing engine
Claude Models
ai-serviceOpus 4.6 for complex reasoning, Sonnet 4.6 for balanced tasks, Haiku 4.5 for fast classification and routing.
Service: Anthropic API
Gemini Models
ai-serviceGemini 3 Pro for multimodal tasks, image generation, and visual understanding.
Service: Google AI API
Local LLM Pool
ai-serviceLlama 4 / Mistral for private queries that shouldn't leave the network. Sovereignty-first option.
Service: Ollama / vLLM
Reasoning Bank
databaseStores decision trajectories, success/failure patterns, and strategy outcomes. Each trajectory includes context, action, result, and quality score.
Service: Supabase + pgvector
Agent Coordinator
computeManages multi-agent workflows. Handles task decomposition, parallel execution, context handoffs, and result aggregation.
Service: Claude Agent SDK
Starlight Vault
storageStrategic memory — long-term patterns, organizational priorities, decision history, and learned preferences. The system's institutional wisdom.
Service: Supabase encrypted storage
Intelligence Telemetry
computeCost tracking, latency monitoring, quality scoring, and model performance comparison across all routes.
Service: Vercel Analytics + custom
Implementation Steps
Gateway & Classification
2 weeks
Build the unified API gateway with intent classification
Tasks
- Deploy SIS Gateway on Vercel Edge Functions
- Train 16-category intent classifier on Claude Haiku
- Implement complexity scoring (1-10 scale)
- Build request normalization pipeline
- Add authentication and rate limiting per tier
Deliverables
Model Router & Multi-Provider
2 weeks
Implement dynamic model selection across providers
Tasks
- Configure Claude API (Opus/Sonnet/Haiku)
- Add Gemini API for multimodal routing
- Set up Ollama for local LLM inference
- Build routing algorithm (complexity × budget × history)
- Implement fallback chains (primary → secondary → tertiary)
Deliverables
Reasoning Bank & Learning
2 weeks
Build the trajectory learning system
Tasks
- Design trajectory schema (context, action, result, score)
- Implement trajectory capture on every request
- Build pattern recognition (which routes work best for which intents)
- Create feedback loop: trajectory → router weights
- Deploy Starlight Vault for long-term strategic memory
Deliverables
Agent Coordination & Telemetry
2 weeks
Multi-agent orchestration and observability
Tasks
- Implement task decomposition for multi-step workflows
- Build context handoff protocol between agents
- Add parallel execution with result aggregation
- Deploy telemetry dashboard (cost, latency, quality)
- Create model performance comparison reports
Deliverables
Code Examples
Dynamic Model Routing with Learning
Routes requests to optimal models based on intent, complexity, budget, and historical performance
type ModelTier = 'opus' | 'sonnet' | 'haiku' | 'gemini-pro' | 'local'
async function routeRequest(req: IntelligenceRequest): Promise<ModelTier> {
const { intent, complexity, budget, userId } = req
// Check reasoning bank for learned patterns
const history = await reasoningBank.getPatterns({
intent,
minScore: 0.8,
limit: 10,
})
// If we have strong historical signal, use it
if (history.length >= 5) {
const bestModel = history
.sort((a, b) => b.qualityScore - a.qualityScore)[0]
return bestModel.modelUsed as ModelTier
}
// Default routing logic
if (req.isPrivate) return 'local'
if (req.hasImages) return 'gemini-pro'
if (complexity >= 8) return 'opus'
if (complexity >= 5) return 'sonnet'
if (budget === 'minimal') return 'haiku'
return 'sonnet' // balanced default
}Cost Estimate
$1,800
per month
$21,600
per year
Assumptions: 200 active users, ~1000 routed requests/day, 30% Opus, 50% Sonnet, 20% Haiku
Use Cases
Technologies
Ready to Build?
Deploy this architecture in minutes, or get the production-ready template with full source code.