Layer Separation

Why: Security, testability, scalability

• Each layer has single responsibility
• Adjacent layer communication only
• Clear protocol boundaries
• Independent deployment

Fine-Tuned Specialists

Why: Domain expertise, cost, speed

• 7B models outperform GPT-4
• 5x faster inference (2-3s)
• $0 cost (self-hosted)
• Privacy compliant

Tool Abstraction (MCP)

Why: Safety, consistency, reusability

• Agents never SSH directly
• Structured JSON responses
• Idempotency built-in
• Multiple agents share tools

Human-in-the-Loop

Why: Trust, compliance, safety

• No auto-execution on prod
• Transparent decision cards
• Audit trail required
• Confidence thresholds

✓ Complexity vs Modularity

More layers = more overhead, but easier to maintain & scale

✓ Latency vs Safety

Human approval adds time, but prevents catastrophic errors

✓ Fine-tuning vs Generic

Training cost upfront, but 10,000x cheaper long-term

✓ Synchronous vs Async

Redis queues enable horizontal scaling & fault tolerance