URI Social

Routing by difficulty, not by default

Sending every request to the strongest model is the obvious way to build this and the reason the economics never work: most posts are not hard, and paying top price for the easy ones is where the budget goes.

SmartLLMRouter grades each request and sends it down one of three tiers: GPT-4o where the work genuinely needs it, Gemini for the middle, and a rule-based path for the cases that were never a language problem in the first place. Cost scales with actual difficulty rather than with worst-case difficulty.

Retrieval that holds up past a demo

Intent analysis runs across Twitter, Facebook and TikTok in real time, over 1,536-dimension embeddings in MongoDB Atlas Vector Search. The search is hybrid: cosine similarity narrowed by metadata filters, because relevance on its own returns the right kind of thing from the wrong account or the wrong month.

Images, and the quality gate in front of them

A multimodal service of around 2,800 lines pairs GPT-4o Vision and DALL·E 3 with ordinary computer vision: blur detection and exposure analysis, behind platform-specific rules for LinkedIn, Instagram, Twitter and Facebook.

The gates matter more than the generation. A model will return something for any prompt, and the question on a brand account is whether what came back is publishable, which is a different question from whether it is plausible.