llm
llmgate
GEPA-for-Router-Policies
Apply GEPA (genetic Pareto reflective optimizer) to LLM router policies — text artifacts with scalar metrics — not just to prompts.
routinggepadspyoptimizationllm-gateway
ANGLES §5. GEPA is the 2026 default prompt optimizer (+13% over MIPROv2, 35x fewer rollouts); nobody applies the same textual-feedback evolutionary loop to router policies. Customers: anyone running a production routing layer that hand-tunes (Portkey/LiteLLM/Kong AI), agent frameworks with multi-model routing (LangGraph, Pydantic AI). Deep-tech: designing a Pareto front over cost/quality/latency/safety GEPA can evolve without exploding sample counts. Kill: DSPy ships dspy.Router with GEPA support.
source: ANGLES.md#5