surogate

Surogate presents a platform for creating AI agents that follow explicit business processes while keeping every piece of data on the infrastructure you choose. The service emphasizes that models are fine‑tuned on your own work history and can run in an EU‑hosted cloud, in your private cloud, on‑premise servers, or in an air‑gapped network. Its flagship model, Surogate 3.7 Flash, is shown side‑by‑side with frontier models such as Kimi K3, GLM 5.2, Claude Opus 4.6, Gemini 3.1 Pro, GPT‑5.4 and DeepSeek V4 Pro, with the site reporting comparable or higher scores on “real agent tasks”.

The platform is built around three main components. First, the desktop client lets you explore an individual agent that works directly on files stored on your computer; the interface shows the agent listening to scanned invoices, extracting fields, filing them into folders, and generating a summary of amounts due. Second, the “Keel” engine orchestrates agents inside well‑defined workflows that can include human approvals, audit logging, and conditional branching. In the Van Dijk Freight demo, an agent receives an incoming invoice, matches it to a purchase order, flags missing PO numbers for review, and updates an invoice log once the user confirms the match. Third, an API key provides programmatic access, allowing developers to call the same agents from other applications or to integrate the AI Gateway, which routes inference requests to multiple providers under a single key.

To fine‑tune a model, you collect chat logs from agents, click a “Build” button, and the system creates a dataset without requiring you to write code. The training command shown on the site (`$ surogate sft config.yaml`) indicates a typical run uses a few million tokens for training and a separate evaluation set, with loss values reported at regular intervals. After training, the model can be served on the same engine that handles inference, meaning you use one environment for both training and production.

Surogate is aimed at enterprises that need high‑throughput LLM inference, strict data residency, and a clear audit trail. It suits organizations in regulated sectors, banks, insurers, government agencies, or any business that wants to automate repetitive document handling, supplier payments, or call‑center case management while keeping data inside its own network.