Skip to main content
Platform

How Divyam.AI Router/Selector Works: Animation

· 4 min read

Divyam.AI runs two loops simultaneously. The first serves your users right now, every request, every second. The second runs in the background in scheduled batches, learning from accumulated traffic to make the first one smarter and cheaper. Here is exactly how both work.

ONLINE LOOP Act and Observe OFFLINE LOOP Learn and Adopt in Batches request deliver consult suggest route response log Your App (originator) Divyam.AI Router Selector v1 · active AI MODELS (100+) GPT-5.5 ChatGPT Gemma 4 Google Gemini Qwen 3.5 Alibaba DeepSeek V3 DeepSeek Log Store online writes · offline reads analytics Dashboard & Billing select logs replay logs responses scores rankings train deploy v2 Experimenter spins up candidates Candidate Model e.g. Llama 5, new release Quality Rubric your quality bar Leaderboard model rankings Selector Trainer trains new selector Selector V2 ready to deploy
Online loop: milliseconds per request  ·  Offline loop: hours per improvement cycle  ·  Selector V2 deployed back into production
The Online Loop: Real Time

Every request from your application hits the Divyam.AI Router. The router consults the Selector, a model trained on your own historical traffic to learn your quality bar and cost constraints, which suggests the best AI model for this specific request. The router forwards the call, receives the response, delivers it back to your app, and simultaneously logs the full interaction to the Log Store. This happens in milliseconds, for every request, continuously.

The Offline Loop: Batch Improvement

In the background, the Experimenter pulls logs from the Log Store and replays them against a Candidate Model, a newly released model like Llama 5 or a cheaper alternative. Each response is scored by the Quality Rubric, your domain-specific quality definition built and maintained in EvalMate, the Divyam.AI product purpose-built to help teams craft rubrics and reward models that mirror how they actually judge their own outputs. The rubric produces a ranked Leaderboard, and the Selector Trainer uses that leaderboard to produce Selector V2, a smarter, updated routing model. V2 is then deployed back into the online loop, replacing the previous selector. The cycle runs again.

For competitive context on which platforms close this loop and which don't: The Infrastructure Gap That Compounds →

For the full platform architecture: The Divyam.AI Platform at a Glance →

Suggested Reading

See both loops in action on your own traffic

Book a Demo