Applied AI R&D: Shipping Production Intelligence Without a Research Lab
Most Australian firms don't need a research lab to do real AI R&D. They need a disciplined loop that turns open questions into shipped systems. Here's the applied approach we use.
"R&D" carries a certain image: a lab, PhDs, a long horizon, and a budget disconnected from any product that ships this quarter. That image is a liability for most companies, because it makes genuine research and development look like something only large institutions can afford. In practice, the R&D that changes a business is rarely frontier research. It is applied — the disciplined work of taking an open question, testing it cheaply, and turning the answer into a system customers actually use.
This is the model MarcTech is built on, and it is worth writing down, because the gap between "we use AI" and "we do AI R&D" is exactly where most of the value sits.
Research is a loop, not a department
The mistake is treating research as a phase that finishes before building begins. In applied work, research and development are the same loop run at different resolutions. You start with a question you cannot answer from documentation alone — can a language model reliably extract structured data from this messy document type? what latency can we actually hit on a live phone call? does retrieval over this corpus stay grounded, or does it drift? — and you resolve it with the smallest experiment that produces a real signal.
The output of a research loop is not a paper. It is a decision: this approach works and here are its limits, or it doesn't and here's why. Every loop either kills a bad idea early or promotes a good one toward production. Run enough of them and you have compounding knowledge about what the current generation of models can and cannot do for your specific problems — which is the only knowledge that matters.
The three questions that gate every build
Before we commit engineering time to an AI feature, three questions have to be answered in order. They are cheap to ask and expensive to skip.
Is the model the right tool at all? A surprising amount of "AI work" is better solved by a query, a rule, or a state machine. Reaching for a model when deterministic code would do is not innovation; it is added cost, latency, and failure surface. The first research question is always whether the problem genuinely needs probabilistic inference.
Off-the-shelf, fine-tuned, or custom? Frontier models are extraordinary general tools, and for many problems the right answer is to prompt one well and move on. But "call the biggest model" is a default, not a strategy. Smaller models are cheaper, faster, and often more than adequate once the problem is scoped tightly. The research question is where on that spectrum the problem actually lives — and that is an empirical question, answered by testing, not by preference.
What breaks it? Every AI system has an edge where it degrades. The research job is to find that edge deliberately, in a test harness, before a customer finds it in production. A feature that works 95% of the time is not 95% done; the last 5% is where the trust is won or lost, and you cannot engineer for it until you have characterised it.
Productionising is the hard part, and it's still research
The industry talks about the leap "from prototype to production" as if it were an engineering formality. It isn't. A demo runs once, under ideal conditions, watched by someone who wants it to work. Production runs thousands of times, under adversarial conditions, watched by no one until it fails.
Closing that gap is itself research. You are answering questions no prototype surfaces: how does the system behave under concurrent load, how do you evaluate quality when there is no single correct answer, how do you observe a probabilistic system well enough to debug it, how do costs scale when usage does. These are not settled problems with textbook answers. They are the applied frontier — and the firms that get good at them are the ones that can ship AI that holds up, not just AI that demos.
Why this is within reach for a small firm
None of the above requires a research institution. It requires three things a focused team can hold: a real problem worth solving, the discipline to test before building, and the engineering depth to productionise honestly. The models are commoditised and improving monthly; the differentiator is no longer access to the technology but the judgement to apply it well — knowing what to build, what to buy, and what to leave alone.
That judgement is accumulated one research loop at a time. It is why applied R&D, done consistently, compounds into a genuine capability rather than a collection of one-off features. And it is why a small Australian firm running tight, honest experiments can ship production intelligence that larger, slower organisations struggle to match.
The lab was never the point. The loop is.