Custom AI systems

AI systems built on your data and wired into your systems

A custom AI system is an assistant or agent that works with your company's own data and tools, under your rules, instead of a chat window that only knows what someone pastes into it. I build them as MCP servers, agents, and internal tools that connect AI clients like Claude to your real systems, and each person only reaches the data they are allowed to see.

Book a callSee the builds
Tools on an existing API
About a day of work
Full system
Usually 2 to 6 weeks
Permissions
Per person, not one shared login
Who builds it
Matthew Ryan, first call to handoff
I

What I build

01

Remote MCP servers

A hosted server that hands any Claude client live, authenticated tools over the network, with nothing to install. Mine runs on Cloudflare Workers with full OAuth, so each person connects their own account and only gets data they are allowed to see.

02

Local MCP toolkits

Tools that run on the machine, so credentials stay on it and nothing sits in the middle. My own twelve-tool toolkit covers the reads, writes, and lookups I do every day.

03

Assistants on your own data

Built on your documents, records, and rules, so answers come from your business instead of the open internet.

04

Agents that do the work

AI that takes an action in your systems, like drafting a follow-up or filing a record. You decide which steps need a person to sign off.

05

Custom machine learning

For when an off-the-shelf model does not fit the job, like the real-time facial landmark model behind Lookzapp.

II

What makes it hold up in production

  • Tools described for the model in plain language. The difference between a tool that gets called correctly and one that gets misused is the description, not the code.
  • Typed inputs checked on every call, so bad input fails loudly instead of quietly returning nonsense.
  • Permissions that follow the person. The AI sees what the signed-in user can see, and nothing more.
  • Hosted where it makes sense, so your team has nothing to install and IT has nothing to push to every machine.
III

About a day, if you already have an API

Once a business has an API, the tools that let an AI client drive it are a day of work, not a project. Most of the effort goes into deciding which actions the AI should be able to take, and describing them well.

IVFit

When this is the right call, and when it is not

A good fit if

  • Your team already uses ChatGPT or Claude and keeps pasting company data into it.
  • Your systems have APIs, and you want AI working with live data instead of stale copies.
  • Different people should see different data, so one shared login will not work.

Probably not a fit if

  • Your team has not used AI day to day yet. Start with training, then build.
  • The real need is a better report or a cleaner spreadsheet. Not everything needs AI.
VCompared

An AI chat app or a custom AI system?

An AI chat app or a custom AI system?
Knows your businessOnly what someone pastes inReads your systems directly, with permission
Takes actionNo. A person copies the answer somewhere else.Yes, through the tools you define
PermissionsWhatever gets pasted in has already left your systemsEach person reaches only their own data
ConsistencyDepends on how each person promptsSame tools and instructions for everyone
SetupMinutesA day for tools on an existing API, weeks for a full system

The straight answer

Start with the off-the-shelf app and train your team to use it well. Built-in connectors cover common tools like shared drives. Build a custom system when people keep copying data between the chat window and your own systems, or when each person should only reach their own data.

VIProof

AI tooling already shipped

MCP

Remote MCP Server

Independent · Cloudflare Workers

A hosted Model Context Protocol server that hands any Claude client a set of authenticated, live tools over the network, with nothing to install locally.

nothing to installremote MCP, OAuth per user

What was built

  • Runs on Cloudflare Workers at the edge, so it is always on, scales on its own, and there is no server to patch or keep alive.
  • Full OAuth flow. Each person connects their own account, and the server only ever returns data that user is allowed to see.
  • Serves a live team leaderboard: the server pulls every member's current numbers, ranks them, and hands the standings back as structured data the model can reason over.
  • Typed tool schemas validated on every call, so the model gets a strict contract and bad input fails loudly instead of quietly returning nonsense.

Impact

Turns a normal chat window into something connected to live company data, with per-user permissions intact and no IT install on anyone's machine.

Model Context ProtocolCloudflare WorkersOAuthEdge runtime
XII

Local MCP Toolkit

Independent · 12 tools

A twelve-tool MCP server that runs on the machine and turns a desktop AI client into something that can actually operate the systems behind it.

12 toolsone desktop client, full reach

What was built

  • Twelve tools covering the reads, writes, and lookups I do daily, each one a single clear action instead of a catch-all.
  • Runs locally, so credentials stay on the machine and there is no third party sitting in the middle.
  • Every tool is described for the model in plain language, which is most of the real work. The difference between a tool that gets called correctly and one that gets misused is the description, not the code.
  • The pattern is portable. Once a business has an API, the tools that let an AI client drive it are a day of work, not a project.
Model Context ProtocolTypeScriptLocal toolingClaude
Lookzapp logo

Lookzapp

Independent · ML and Web

lookzapp.com

A consumer web app built on TensorFlow that maps 168 landmark points across a face in real time to compute structural metrics, then returns a scored analysis.

168facial landmark points, live

What was built

  • Trained and deployed a custom ML model for real-time facial landmark detection.
  • Built a real-time camera pipeline mapping 168 individual landmark points per frame.
  • Computed golden-ratio alignment, canthal tilt, and facial symmetry, returned as a scored analysis with supporting data.
TensorFlowCustom ML modelReal-time camera pipelineWeb
VIIProcess

How a build actually goes.

The same five steps on every engagement, so you always know what happens next and what it costs before it happens.

  1. I

    Discovery call

    Thirty minutes on how the business actually runs. What the team touches by hand, what breaks, and what nobody has time to fix. No slides and no pitch deck.

  2. II

    Process mapping

    I sit with the people doing the work and write down what actually happens, step by step. The real bottleneck is usually somewhere nobody flagged on the call.

  3. III

    Proposal with hours

    A fixed scope. Every piece listed with the hours it takes and what it costs. You approve the number before anything gets built.

  4. IV

    Build with approval gates

    Work ships in pieces and you sign off on each one. No multi-month black box, and no finding out at the end that it solves the wrong problem.

  5. V

    Handoff

    You get the code, the accounts, and written documentation covering how it works, what it touches, and what to do when something looks wrong. It runs without me.

  6. Optional

    Then, if you want it, a retainer

    Most clients keep me on a flat monthly retainer after handoff for monitoring, fixes, and small changes. It is optional, and everything is built so you are not stranded if you skip it.

VIIIQuestions

Questions about custom AI systems

What is an MCP server?

MCP, the Model Context Protocol, is a standard way to give an AI client like Claude a set of tools it can call, like looking up a customer or creating an invoice. An MCP server is the piece that exposes your systems as those tools. It can run locally on one machine, or be hosted so a whole team connects with nothing to install.

Is our data safe?

The system is built so the AI only reaches data the signed-in person is allowed to see. Hosted servers use OAuth, so each person connects their own account, and local tools keep credentials on the machine. What a tool returns goes to the AI provider you use, so which provider, and under what terms, gets decided with you before the build starts.

Which AI models do you work with?

Most of my tooling is built for Claude, and MCP also works with other clients that support the protocol.

How long does it take?

Tools for a system that already has an API can be a day of work. A full assistant or agent usually lands in the same two to six weeks as most first builds, depending on how many systems it touches.

IXRelated

Often built alongside this

AI trainingLive sessions that turn an AI account into a real part of how your team works. No technical background needed.See the serviceBusiness process automationThe repetitive work sitting between your systems, handled end to end. Nobody keys the same data in twice.See the serviceCustom appsWeb and iOS apps, websites, and portals built around the workflow your team already has.See the service

Have a process that should run itself?

Tell me how your business actually works day to day, and I will show you what can be automated away. Book a time, or call or text me directly.

Book a callCall or text · 716 263 2627matt@braceyskyway.com