AI automation and LLM features
A language model is useful when a task involves reading or writing: understanding an email, drafting a reply, answering a question from your own documents. I add that step to real workflows and build the controls around it.
Email me about ai automation ↗Who it is for
- A business that wants AI to do a specific job, not a general chatbot
- A team with documents or a website that people keep asking the same questions about
- A product that needs an AI feature built by someone who will also make it reliable
What it covers
LLM APIs called from Python, C# or TypeScript, connected to email and the tools you already use. Where a workflow platform fits better, I use one.
- Reading and classifying incoming messages
- Drafting replies and outreach from approved wording
- Assistants that answer only from your own content and show their sources
- Limits on usage and spending, and an off switch
- Tests for wrong answers and for people trying to misuse it
How a project runs
- 01
Pick one job
We choose a single task where a wrong answer is cheap to catch, and decide what the model may and may not do.
- 02
Ground it
The model works from your approved content or templates, not from whatever it remembers.
- 03
Put controls around it
Checks on what it produces, limits on how much it can be used and spent, and a person in the loop where it matters.
- 04
Test, then release
It is tested with real questions and deliberately awkward ones before anyone depends on it.
Work that shows it
This site’s assistant is built this way: it retrieves from reviewed content, checks every quoted claim against its source, and runs inside a spending limit. It stays switched off until its evaluation is complete.
Read the common questions ↗Questions
01Is this a chatbot?
It can be, but it does not have to be. Often the useful part is one step inside a workflow, such as reading an email and drafting the reply for someone to approve.
02How do you stop it making things up?
By limiting it to your approved content, checking what it produces against that content before it is shown or sent, and keeping a person in the loop where a mistake would matter.
03What will it cost to run?
Model providers charge by usage. I estimate that before building and add limits so the bill cannot run away.