NOMINAL The optimal inference path for AI requests

[Problem]: AI applications often send every request either to a premium model, increasing cost and latency, or to a cheaper model, risking quality. But AI workloads have different reasoning, coding, context, quality, and latency requirements. [Solution ]: NOMINAL is an OpenAI-compatible AI inference control plane that dynamically finds the lowest-cost model predicted to satisfy each request. It profiles incoming conversations across multiple capability dimensions, evaluates candidate models based on capabilities, cost, and latency, and selects the optimal inference path. After execution, NOMINAL evaluates response quality. If the result misses the required quality target, it can escalate to a higher-capability model. Separately, if a provider times out, rate-limits, or becomes unavailable, NOMINAL performs infrastructure fallback to another valid candidate. [Key Innovation]: NOMINAL separates quality escalation from infrastructure fallback, making inference routing explainable, bounded, resilient, and auditable rather than relying on a simple “easy vs. hard” classifier. Infrastructure Impact Every decision is observable. NOMINAL tracks request profiles, candidate models, routing reasons, model paths, tokens, configured-price cost, latency, quality, escalation, fallback, and status. Its dashboard visualizes infrastructure efficiency and request-level decision traces. A reproducible 40-prompt benchmark compares Always Frontier, Always Economy, and NOMINAL using real provider responses and token usage, without claiming unverified savings. [Features & Stack]: OpenAI-compatible gateway, capability-aware routing, OpenAI/Ollama adapters, quality gates, escalation/fallback, SQLite telemetry, benchmark engine, React dashboard, Docker Compose, health checks, and safe error handling. Built with Python, FastAPI, Pydantic, httpx, SQLAlchemy, SQLite, React, TypeScript, Tailwind CSS, Recharts, Docker, and Nginx.