Multi-Model Admin Agent

The problem

Choosing a model on reputation is guesswork. Comparing models fairly means asking them the same question over the same retrieved evidence — otherwise two answers can differ simply because they read different documents.

What I built

• One model picker across Azure OpenAI, Microsoft Foundry, Anthropic Claude and Google Gemini, with a health check and per-model reasoning-effort settings.
• Compare mode: two to four models answer side by side from one search and one prompt, with time and cost per answer.
• Optional review by a second model that checks a draft against the evidence — a model never reviews its own draft.
• Server-side web search and page reading behind a guard that refuses internal addresses and re-checks every redirect, plus document attachments that are read and never stored.
• Streaming answers across four different provider APIs; token usage is counted even on failed turns, and an unknown price is reported as unknown, never as zero.