RESEARCH

Intelligence Delivery Network

CDN-style routing model for LLM workloads

route requests to the right compute tier before spending frontier-model money

What it is

Intelligence Delivery Network, or IDN, is a CDN-style architecture for AI workloads. Instead of sending every request to the same large model, IDN profiles the request first and routes it to the compute tier that fits its complexity, privacy needs, latency target, and tool requirements.

The central idea is simple: trivial, private, or latency-sensitive requests should stay close to the user; expensive global frontier models should be reserved for work that actually needs them.

The tier model

Tier Location Typical use
L0 On-device Offline use, privacy-forced prompts, PII preprocessing
L1 Edge Fast extraction, classification, lightweight RAG
L2 Regional Coding, summarization, document workflows, tool agents
L3 Global Deep reasoning, research synthesis, multi-agent planning

Core components

Workload profiler

Analyzes token size, complexity, reasoning hops, domain, PII risk, and latency sensitivity.

Router

Maps profiler output to a tier, expert model, tool plan, fallback path, and cost estimate.

L0 privacy firewall

Keeps data on-device when egress is blocked and can redact sensitive fields before escalation.

Async planner

Splits multi-part requests into subtasks that can route to different tiers and merge cleanly.

Status

Development is currently on hold. The repository captures the architecture, routing model, profiler schema, and prototype direction.

See the GitHub repo for the current design notes.