Architecture Overview

Meida operates as an edge-distributed routing gateway positioned between your application and multiple foundation LLMs. Every turn passes through a 3-step pipeline before reaching a model.


Routing Pipeline

01CLASSIFY

Semantic Classification

Each incoming turn is analyzed in under 4ms by a lightweight local classifier. It extracts intent, complexity score (0–1), context length, and risk flags — without ever touching a foundation model.

Latency
<4ms
Model
Local Rules + MicroLM
Output
Complexity Score + Intent Tag
02ARBITRAGE

Tier Selection & Cost Arbitrage

Based on the classification output, the arbitrage engine selects the lowest-cost provider tier that can reliably complete the task. Simple queries go to fast/cheap models; complex reasoning escalates to premium tiers.

Latency
<8ms
Model
Small Judge (Llama-3 8B)
Output
Tier Level (T1–T4) + Provider
03FAILOVER

Zero-Downtime Failover

If the primary provider returns a 429 (rate limit), 503 (outage), or times out, the failover engine instantly reroutes to the next-best provider in the same tier. No tokens are lost, no latency spikes — the user never notices.

Failover Time
<12ms
Token Loss
0.00%
Fallback Depth
3 Providers Deep

Edge Node Pool

The following providers are actively connected to the Meida routing mesh. Each node is health-checked every 5 seconds.

ProviderTierRoleStatus
Claude 3.5 SonnetT4 (Premium)Complex reasoning, audits, long-contextPRIMARY
GPT-4oT3 (Standard)General purpose, balanced cost/qualityACTIVE
DeepSeek-V3T2 (Fast)Code generation, structured outputHOT
Llama-3 70BT1 (Economy)Simple queries, classification, chatHOT
End-to-End Routing Overhead
<50ms total

Classify (4ms) + Arbitrage (8ms) + Network Hop (~30ms) = sub-50ms before first token.