Custom LLMs

Large Language Models trained specifically on your business data, delivering responses that truly understand your company, industry, and customers.

Why Custom LLMs?

Your Data, Your Model

Trained exclusively on your business knowledge, processes, and industry expertise

Secure & Private

Define the hosting boundary and any external model calls before deployment; verify the agreed data flow

Industry-Specific

Understands your sector's terminology, regulations, and best practices

Compliance Ready

Built to meet Australian data protection and industry compliance requirements

Your Knowledge

Private & Secure

Perfect for Your Business

Unlike generic AI models, our Custom LLMs are trained exclusively on your business data, ensuring every response aligns with your brand, policies, and expertise.

Customer service with your exact tone and policies

Technical documentation and knowledge base queries

Compliance checking and regulatory guidance

Internal training and onboarding assistance

Product recommendations based on your catalog

Industry-specific analysis and reporting

Key Advantages

No Generic Responses

Every answer is tailored to your business context and standards

Faster Implementation

Pre-trained on your data means immediate, accurate responses

Competitive Advantage

Your unique knowledge becomes a powerful business asset

Our Custom LLM Process

Data Analysis

We analyze your existing knowledge base, documents, and processes

Custom Training

Train the model on your specific data and business context

Testing & Refinement

Rigorous testing to ensure accuracy and brand alignment

Deployment

Secure deployment in your environment with ongoing support

Four Ways to Build This, and How We Choose

"Custom LLM" covers several quite different technical approaches, and the difference between them is the difference between a project that costs a few thousand dollars and one that costs a great deal more. Part of the discovery work is deciding which of these you actually need, and we are happy for that answer to be the cheapest one.

Retrieval over your documents

Best for: Most Australian SMEs, most of the time

Your policies, manuals, price lists, past quotes and support history are indexed and searched at the moment a question is asked, and the model answers using what it retrieved. Nothing is baked into the model, so correcting an answer is a matter of correcting the source document. This is the approach we recommend first, because it is cheaper to run, far easier to keep current, and it can cite where each answer came from.

Fine tuning on your material

Best for: Consistent tone, format and classification work

A base model is further trained on your own examples so it adopts your house style, your formatting conventions and your way of categorising things. Fine tuning teaches behaviour well and facts poorly, so we use it for how something should be said rather than what should be said, and usually alongside retrieval rather than instead of it.

Self hosted deployment

Best for: Data that genuinely cannot leave your environment

An open weights model runs on infrastructure you control, whether that is your own server or a private cloud tenancy. This is the right answer where a contract, a regulator or a client agreement requires it. It costs more to run and maintain than a hosted model, so we will tell you honestly when your obligations do not actually require it.

A guarded prompt layer

Best for: Getting value quickly from a well defined task

Sometimes the honest answer is that you do not need a custom model at all. A carefully constructed instruction layer over a capable general model, with your data supplied at the point of the question and hard limits on what it may say, solves a surprising share of real business problems in a fraction of the time. We would rather build you that than sell you a training project you did not need.

What This Solves in Practice

The businesses that get the most out of this are the ones sitting on years of accumulated knowledge that only exists in documents and in a few people's heads. Four situations we see regularly in Australian firms.

A construction firm with fifteen years of past quotes

How it works today

Estimators dig through old job folders to find how a similar scope was priced last time, and the knowledge walks out the door when a senior estimator retires.

With a custom model

The team asks a question in plain English and gets the relevant past jobs, the assumptions used and the exclusions that were applied, with links to the original documents so the estimator can check the source before relying on it.

A member association fielding the same policy questions

How it works today

Members email questions that are answered in a 90 page handbook nobody reads, and staff answer them one at a time with slightly different wording.

With a custom model

A model grounded in the current handbook answers in your language and your policy terms, quotes the clause it relied on, and refers anything involving a dispute or an exemption to a person.

A distributor with a complex product catalogue

How it works today

Sales staff cannot keep three thousand product variants and their compatibility rules in their heads, so quotes go out with the wrong part and get corrected later.

With a custom model

Staff describe the application and the model surfaces compatible products with the specification that supports it, so the check happens before the quote goes out rather than after.

A professional services firm onboarding new staff

How it works today

New hires spend their first months interrupting senior colleagues about internal process, and the answers depend on who is asked.

With a custom model

An internal assistant trained on your actual documented processes answers the routine questions consistently, and flags the ones where your documentation genuinely does not cover the case.

Honest Limits Worth Knowing First

Language models are genuinely useful and they are not magic. Understanding where they struggle before you commit budget is the difference between a tool your staff rely on and one they abandon after a fortnight.

Arithmetic and reconciliation

A language model is the wrong instrument for adding up a ledger or reconciling a bank feed. Those jobs belong in software that calculates deterministically. Where a task mixes the two, we have the model read and route while the numbers are computed properly elsewhere.

Knowledge nobody ever wrote down

If the reason your senior estimator is good is instinct built over twenty years and never recorded anywhere, there is nothing for the system to learn from. Capturing that knowledge is a worthwhile project in its own right, and it comes before the model, not after it.

Contradictory source material

Where three versions of a policy exist across a shared drive and none is marked current, the system will answer from whichever it retrieves. Deciding which document is authoritative is a business decision, and one of the more valuable side effects of doing this work at all.

Anything requiring accountability

Regulatory advice, clinical guidance and legal interpretation need a qualified person who is answerable for the answer. We build these systems to recognise those questions and route them to a human rather than attempt them.

What It Costs

Builds are scoped and quoted at a fixed price after discovery. There is no useful sticker price for a custom model, because the work is dominated by the state of your source material. A firm with well maintained documentation and a firm with fifteen years of unsorted PDFs are very different projects even when they want the same outcome.

The consulting engagement that produces the scope and the fixed build price is $1,000 for a focused piece of work, $3,000 for a standard engagement and $10,000 for a broader strategic engagement. Initial consultations are free for businesses with 20 or more full-time staff (smaller teams pay $200 including GST).

Running costs are separate and worth understanding before you commit. Hosted models are billed on usage, which is usually modest for internal tools; a self hosted deployment trades that for infrastructure you pay for whether it is busy or idle.

How Long It Takes

A retrieval based assistant over a defined set of documents is typically in the hands of a pilot group within four to eight weeks. That includes the part people forget: testing it against the questions your staff actually ask, rather than the questions that make a demonstration look good.

Fine tuning adds time for assembling and reviewing training examples. Self hosting adds infrastructure setup and hardening. Document clean-up is the most common reason a timeline stretches, and it is also the work that most improves the result.

We publish work under our own name rather than invented case studies. Reignite Health, Firebox Australia and Nutrition Science Group are three Australian engagements you can ask us about directly.

Frequently Asked Questions

Do we really need a custom model, or is a general one enough?

Often a general model with your own documents supplied at the moment of the question is enough, and it is faster and cheaper to get there. A custom build earns its place when your terminology, your rules or your privacy obligations mean a general model keeps giving answers that are close but wrong. We will tell you which situation you are in before you spend anything on training.

Where does our data actually go?

That depends on the deployment you choose, and it is a decision we make with you rather than for you. Options run from a hosted model with contractual commitments on data handling, through to a model running entirely on infrastructure you control where nothing leaves your environment. We document exactly which data travels where as part of the build, and where personal information is involved we design in line with the Privacy Act 1988 and the Australian Privacy Principles.

How much of our data do we need before this is worth doing?

For the retrieval approach, less than most people expect. If your policies, procedures and past work are written down somewhere, even messily across folders and email, that is usually enough to start. Fine tuning is hungrier and wants hundreds of good examples of the behaviour you want, which is one of several reasons we rarely start there.

What stops it from making things up?

Grounding and boundaries, not hope. Answers are drawn from your retrieved source material and cite it, so a claim with no source behind it is visible as such. We define the topics the system must refuse and route to a human, we test it against questions designed to trip it, and we keep a log so a wrong answer can be traced back to the document that caused it.

What happens when our policies change?

With retrieval, you update the document and the next answer reflects it. That is the main practical reason we favour it over training your knowledge into a model, where a policy change means retraining. Keeping the source material current becomes an ordinary business task rather than an AI project.

What does a custom LLM cost?

Every build is scoped and fixed-priced after discovery, because cost turns almost entirely on how much of your knowledge is already written down and how many systems the assistant needs to reach. The consulting engagement that produces that scope is priced clearly: $1,000 for a focused piece of work, $3,000 for a standard engagement and $10,000 for a broader strategic one. Initial consultations are free for businesses with 20 or more full-time staff (smaller teams pay $200 including GST).

How long does it take?

A working retrieval assistant over a defined body of documents is usually live for a pilot group within four to eight weeks. Builds involving self hosting, significant document clean-up, or connections into several internal systems run longer, and we will give you a realistic range after discovery rather than an optimistic one before it.

Ready for Your Custom LLM?

Transform your business knowledge into a powerful AI assistant.

Book a Consultation