Skip to content
Cheatpot
AI Lab

Where we turn frontier AI into production systems.

Agents, retrieval, custom models and MCP servers — built with evaluation and guardrails from day one, and connected to the systems your business already runs on.

Start an AI project Explore MCP
Capabilities

What the lab builds.

Autonomous & assistive agents

Agents that plan, call tools and complete multi-step work — with human approval where it matters.

  • Tool use
  • Planning
  • Human-in-the-loop

MCP servers

Model Context Protocol servers that expose your APIs, databases and actions to AI clients — securely and with full audit trails.

  • Tools
  • Resources
  • OAuth

Retrieval & knowledge

Assistants grounded in your documents and data, with citations and access control respected end to end.

  • RAG
  • Vector search
  • Hybrid search

Custom & fine-tuned models

Domain-adapted models for classification, extraction and generation where general models fall short.

  • Fine-tuning
  • Embeddings
  • Distillation

AI-powered automation

Document intelligence, triage, summarisation and data entry pipelines that remove repetitive work.

  • OCR
  • Extraction
  • Workflows

Evaluation & guardrails

Test sets, automated evals, monitoring and cost controls so AI quality is measured — not assumed.

  • Evals
  • Observability
  • Safety
Model Context Protocol

The open standard that plugs AI into your systems.

MCP lets AI assistants discover and use tools, read resources and follow prompts exposed by a server you control. We design MCP servers around your domain — scoped, authenticated and observable — so any compatible AI client can work with your business safely.

  • Tools for real actions
  • Resources for live context
  • OAuth-secured remote servers
  • Audit logs for every call
orders-server.tsTypeScript · MCP SDK
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";

const server = new McpServer({ name: "orders", version: "1.0.0" });

server.registerTool(
  "get_order_status",
  {
    description: "Look up the live status of a customer order",
    inputSchema: { orderId: z.string() },
  },
  async ({ orderId }, { authInfo }) => {
    await assertScope(authInfo, "orders:read");
    const order = await db.orders.find(orderId);
    return {
      content: [{ type: "text", text: `${order.id}: ${order.status}` }],
    };
  },
);
Illustrative example: a scoped, authenticated MCP tool exposing order status to AI clients.
Method

Evaluation-driven, from idea to operation.

  1. 01

    Use case

    Define the job and the metric

  2. 02

    Eval set

    Real examples, expected outputs

  3. 03

    Prototype

    Models, prompts, tools, retrieval

  4. 04

    Harden

    Guardrails, auth, latency, cost

  5. 05

    Operate

    Monitoring and continuous evals

Principles

How we build AI responsibly.

Evaluate first

Every system starts with a measurable definition of 'good' and an evaluation set to prove it.

Least privilege

AI gets exactly the data and actions it needs — scoped, authenticated and logged.

Model-agnostic

We pick the right model for quality, latency, cost and data-residency needs, and keep you portable.

Humans in control

High-impact actions stay reviewable, reversible and explainable.

Have a workflow AI should be doing?

Bring us the process. We'll tell you honestly whether AI can do it well — and build it if it can.

Talk to the AI Lab
hello@cheatpot.com