AI features and integrations
GPT-powered classification, search and assistants, added to the product you already have.
The problem
Your users keep asking for AI and your team has no time to research it. You are not sure which model to use, what it will cost to run, or how it will fail. You want one feature that works, not a science project.
What you get
- 01An AI feature inside your existing product: classification, search and reranking, or an assistant
- 02Model choice backed by tests on your data, including open-weight options
- 03An evaluation set so you can see accuracy, not guess it
- 04Clean API integration with your current backend
- 05Fallbacks for when the model is wrong or slow
How it runs
- 1Pick one feature with a clear success measure
- 2Test candidate models on a sample of your real data
- 3Build the feature behind your existing API, with a demo every week
- 4Ship, measure, tune
Proof
Questions
Can you add GPT to my existing app?
Yes. I add LLM features through your existing backend and API, so the rest of your app does not change. For Margintall I built a GPT-based transaction categorisation engine in Python.
How do I use AI to categorise data automatically?
Send each item with clear category definitions to an LLM, validate the output against a fixed list, and keep a labelled test set to measure accuracy. I built a GPT-based transaction categorisation engine for Margintall, and this is how I approach new classification features.
Should I use OpenAI or an open-source model?
It depends on your data, cost and privacy needs, so I test both on your data before deciding. I co-authored a research paper that stress-tested open-weight LLMs, including OpenAI's gpt-oss, as recommendation rerankers. It was accepted at the SIGIR 2026 AgentSearch workshop and ranked joint-1st by reviewer score.
Can an LLM improve search or recommendations on my site?
Often, as a reranker on top of your current search. LLMs also fail in specific, repeatable ways, which is what my SIGIR 2026 workshop paper measured. I build in checks for those failures.
How long does it take to add an AI feature?
A single, well-defined feature usually goes from test to working demo in a few weeks. You see progress in a weekly demo.
Other services
04 / Contact
Have something to build? Book a call.
20 minutes, free. Tell me what you need; you leave with a scope, a timeline and a straight answer on whether I'm the right fit.