Where are Curtly proxy nodes located?

Proxy nodes are deployed across North America, Europe, Asia, and Latin America closest to major cloud provider AI clusters.

Edge Performance Engine

Sub-15ms ultra-low latency gateway governor

Engineered for high-frequency trading bots, real-time voice assistants, and interactive coding IDEs where every millisecond counts.

P50 Processing Latency

< 4ms

In-memory deterministic compression execution.

P99 Edge Latency

< 15ms

Guaranteed overhead under high concurrent load.

Streaming TTFT

Instant

Zero buffer accumulation during token generation.

The production engineering challenge

Adding middleware to AI pipelines often introduces unacceptable 200-500ms latency, causing sluggish IDE autocompletions and broken voice agent conversational flow.

Technical architecture & mechanics

Global edge runtime deployment

Executed on edge servers in 300+ cities worldwide, routing requests to the nearest AI model region.

Zero cold-start in-memory tokenization

BPE vocabulary hash tables are pre-warmed in memory at boot time, eliminating initial request lag.

Direct stream pass-through pipeline

Streams response chunks immediately to the client with sub-millisecond SSE forwarding.

Real-Time Streaming Example

import OpenAI from 'openai'; const client = new OpenAI({ baseURL: 'https://proxy.curtly.dev/v1', apiKey: process.env.OPENAI_API_KEY, }); const stream = await client.chat.completions.create({ model: 'gpt-6-astra', messages: [{ role: 'user', content: 'Stream code solution...' }], stream: true, }); for await (const chunk of stream) { process.stdout.write(chunk.choices[0]?.delta?.content || ''); }

Architecture comparison

Evaluation metric	Curtly.dev gateway	Traditional / naive approach
Proxy Overhead	Sub-15ms global edge processing	150-500ms centralized server hops
Streaming Protocol	Direct zero-copy Server-Sent Events (SSE)	Full response buffering before forwarding
Cold Start Behavior	0ms (pre-warmed in-memory BPE tables)	1-3 second container initialization delays

Frequently asked questions

No. Input compression runs in under 4ms before dispatch, and output efficiency reduces total tokens so streaming actually finishes faster overall.

Explore related platform features

Core Engine Architecture

Ready to optimize your AI infrastructure?

Start optimizing in 60 seconds with a 1-line baseURL change. Free tier available with zero credit card required.

Start optimizing free Simulate token ROI