How a custom LLM actually works
No jargon and no hype — what "custom AI" really means, the three ways it gets built, which one your situation calls for, and what our process looks like from first conversation to a working system.
What "custom" actually means
A large language model is a general-purpose text engine. It's trained on a broad slice of the internet, it's good at language in general, and it knows nothing at all about your business, your customers, or your records.
"Custom" almost never means training a model from scratch — that costs millions and is rarely the right answer. It means taking a strong existing open-weight model and connecting it to your world in one of the three ways below, so that its answers are grounded in what's actually true about your business rather than what sounds plausible.
Three ways to make a model yours
Most real projects use a mix of these rather than just one. Which combination depends entirely on the job you need done.
Prompting
You give the model instructions and examples directly in each conversation. No setup, works immediately, and resets every time — nothing is remembered between sessions.
RAG
The model searches your actual documents or database first, then answers using only what it found — and tells you which document the answer came from.
Fine-tuning
The model is trained further on examples from your domain, so your tone, format, and patterns are built into it rather than explained each time.
So which one do you need?
In practice the decision is usually this simple.
What a RAG pipeline is actually doing
This is the most common setup for a business that wants AI answering from its own records. Five steps, and the last one is the only part anybody sees.
Pipeline log
Why it's built this way
How AIXFIN builds it
The same sequence on every project, whether it's a small intake assistant or a full private deployment.
Understand the job
We map exactly what the system needs to answer or automate. Not a general-purpose assistant — one specific job, done well enough to rely on.
Choose the approach
Prompting, RAG, fine-tuning, or a mix — scoped to what the task actually needs rather than the most impressive option available.
Choose where it runs
A managed API when speed matters most, or a self-hosted model when your data can't leave your systems. Your call; we build for either.
Connect your data
Documents, records, and past jobs get cleaned, structured, and indexed so the model retrieves from them accurately instead of approximately.
Test against real cases
We run it against real questions and real edge cases from your business — including the awkward ones — before anyone else touches it.
Ship and hand over control
You get a working system, visibility into how it reaches its answers, and the ability to adjust it. Not a black box you have to trust blindly.
Common questions
Do I need my own servers to run a local model?
Not necessarily. "Local" means infrastructure you control, which can be a modest cloud instance rather than hardware in your office. We scope it to your budget and to how sensitive the data actually is.
Is RAG the same thing as fine-tuning?
No. RAG looks things up at the moment of answering; fine-tuning changes the model itself in advance. RAG is far easier to keep current — you just update the document. Fine-tuning is better when you need a consistent style or behaviour baked in.
How long does a build like this take?
A focused RAG system over an existing set of documents can be a matter of days. Fine-tuning or a full private deployment takes longer, and how much longer depends almost entirely on how ready your data is.
What if the model gets something wrong?
It will, occasionally — anyone claiming otherwise is selling something. What matters is that you can see why. A RAG system shows its sources, so a wrong answer traces back to a specific document you can fix, rather than disappearing into the model.
Have a specific use case in mind?
Tell us what you want the system to answer or automate, and we'll tell you which of the three approaches actually fits — including when the answer is "none of them yet."