Skip to content
Let’s talk

All technology

The right model for each task, running where your data is allowed to go

We choose models per task, not per project. A small, fast model classifies and extracts; a frontier model handles reasoning and writing. Every call goes through one gateway that routes requests, enforces budgets and logs cost.

Example: one gateway routing to two modelsYour application sends every model request to a model gateway. The gateway routes each request either to Claude through your own provider account or to a Llama model self-hosted on vLLM, and writes every call to a budget and cost log.Your applicationModel gatewayClaude, youraccountLlama on vLLMBudgets, cost log

When clients need this

  • Model costs are growing faster than usage, and nobody can say which feature is responsible.
  • Some data can’t be sent to a third-party API, but you still want to use language models on it.
  • A single provider outage takes your AI features down.

How we build it

Closed-weight models such as Anthropic Claude, OpenAI GPT and Google Gemini are our default where quality matters most. We use them through your own provider accounts, under your data agreements, so the bill and the data terms are yours.

Open-weight models such as Llama, Qwen, Mistral and DeepSeek run on a hosted inference provider, or self-hosted with vLLM in your cloud, when cost, data residency or on-premise requirements call for it.

The gateway enforces a key per application, monthly budgets with hard caps, alerts at 50, 80 and 100 percent of budget, prompt caching, and cost logging on every request. If one provider fails, requests fall back to a second.

Token costs grow with how much your AI features are used, not with website traffic.

daily cost = requests × (input tokens × input price + output tokens × output price)

example: 1,000 conversations a day × (2,000 input + 500 output tokens)
       = 2M input tokens + 0.5M output tokens a day
How we estimate token cost before building. Prices come from the provider’s current pricing page.

What we don’t do

  • We don’t lock you into one model provider.
  • We don’t claim your data never leaves your servers when a third-party model API is in use. We tell you which applies, system by system.
  • We don’t run models in production without budgets and alerts.