Intelligent Routing System (IRS)

An adaptive edge-cloud AI routing system that intelligently routes coding queries between local (Qwen Coder) and cloud (Claude Code) models based on task complexity.

Quick Start

# Install dependencies
python3.11 -m pip install -r requirements.txt

# Set up environment variables
cp .env.example .env
# Edit .env and add your credentials (POSTGRES_PASSWORD, GROQ_API_KEY)

# Interactive mode
./run.sh generate --interactive

# Health check
./run.sh check

# View stats and performance
./run.sh stats
./run.sh dashboard

# Evaluate routing system
./run.sh evaluate data/datasets/dataset_test.jsonl

# Compare routing strategies
./run.sh compare data/datasets/dataset_test.jsonl

Documentation

• PROGRESS.md - Complete progress report with all phases, features, and running instructions
• IMPLEMENTATION_PLAN.md - Full 7-phase project roadmap

Current Status

Phase 1: ✅ Complete

• Intelligent routing between local and cloud models
• Conversation continuity support
• Rich CLI interface with statistics
• Performance tracking and logging

Phase 2: ✅ Complete

• Advanced feature extraction (19 features + embeddings)
• Semantic embeddings (384-dim using sentence-transformers)
• Complexity signal detection (concurrency, algorithms, reasoning)
• PostgreSQL + pgvector for similarity search (hosted on Supabase)
• Enhanced routing strategy

Phase 3: ✅ Complete

• Dataset generation infrastructure
• Dual-model evaluation (Qwen + Claude)
• 100+ seed prompts across 10 categories
• LLM-as-judge evaluation using Groq (llama-3.1-8b-instant)
• Train/val/test split generation
• JSONL dataset format
• Automated dataset generation and training pipeline

Phase 4: ✅ Complete

• Neural network routing model (145K parameters)
• 3-layer feedforward architecture with dropout and batch normalization
• 86.67% test accuracy on routing decisions
• ML-based routing (now default)
• Enhanced rule-based routing (fallback)
• Configurable decision threshold
• Model persistence and loading

Phase 5: ✅ Complete

• Multi-objective cost optimization (accuracy, latency, cost)
• Confidence-based routing with automatic escalation
• Threshold tuning and optimization
• Strategy comparison framework

Phase 6: ✅ Complete

• Comprehensive evaluation framework
• Dataset evaluation with test files
• Live model comparison
• Performance dashboard CLI command
• Automated report generation
• 66.7% cost reduction vs always-cloud baseline

---

For detailed setup, usage, examples, and complete progress report, see PROGRESS.md