One private endpoint for every open model you run.

import os
from openai import OpenAI

client = OpenAI(
    base_url=os.environ.get("ESIN_GATEWAY_URL", "https://ai.esinai.com/api"),
    api_key=os.environ["ESIN_API_KEY"],             # never hard-code keys
)

raw = client.chat.completions.with_raw_response.create(
    model="auto",                                     # let the gateway route
    max_tokens=800,                                   # bound every request
    messages=[{"role": "user", "content": "Q3 satış trendlerini analiz et."}],
    extra_headers={"x-data-class": "internal"},      # drives routing and residency
)
print(raw.headers.get("x-esin-route"))               # which model, which rule
print(raw.parse().choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: process.env.ESIN_GATEWAY_URL ?? "https://ai.esinai.com/api",
  apiKey: process.env.ESIN_API_KEY,                 // never hard-code keys
});

const { data, response } = await client.chat.completions
  .create(
    {
      model: "auto",
      max_tokens: 800,
      messages: [{ role: "user", content: "Analiseer de Q3-verkooptrends." }],
    },
    { headers: { "x-data-class": "internal" } },
  )
  .withResponse();

console.log(response.headers.get("x-esin-route"));
console.log(data.choices[0].message.content);

curl -sS -D - "$ESIN_GATEWAY_URL/chat/completions" \
  -H "Authorization: Bearer $ESIN_API_KEY" \
  -H "Content-Type: application/json" \
  -H "x-data-class: internal" \
  -d '{"model": "auto", "max_tokens": 800,
       "messages": [{"role": "user", "content": "Analyse Q3 sales trends."}]}'
# -D - prints response headers, including x-esin-route

# Inference node: listen on the private interface only.
# The gateway owns the public edge, authentication and quotas.
vllm serve openai/gpt-oss-120b \
  --host 10.0.0.12 --port 8000 \
  --api-key "$VLLM_INTERNAL_KEY" \
  --max-model-len 65536 \
  --gpu-memory-utilization 0.90

# Firewall: allow port 8000 from the gateway host only.
# Pin the vLLM version and the model revision in your deployment manifest.