Frugal Yet Powerful: Under the Hood of m19's AI Engine

Everyone is selling AI these days. The issue in this ubiquitous claim is that AI means everything and nothing all at once. At m19 we, too, have our AI. Or should we say our "intelligent solution." In an age where everyone is battling over token counts, parameter sizes, memory throughput or query latency, we stay true to our core missions: we crush the complexity of running Amazon ad campaigns and we optimize for profit. Or to be more accurate, "your profits." All while staying true to our core values: technical excellence, ultra efficiency, zero bu**shit.

In this piece, we will invite you backstage. Your guide will be our CTO, Pascal Pons, a former employee at the Criteo Predictive Search lab. This PhD in computer science is a firm believer that less is more. And that the frugality of a system is proof of its efficiency.

But before digging into the engine, we must start with the constraints. In our case, Amazon data.

‍

The Data Limitation: Why Amazon Doesn't Make This Easy?

"The limit of our craft is not in the algorithm: it's in available data. Amazon does not natively provide statistics at the level of granularity that would make a machine learning model truly powerful. It does not give you performance statistics broken down by both keyword and ASIN simultaneously," states Pascal Pons. This is his way of explaining that data couldn't be the starting point — as it is for so many AI projects, like LLMs. So, we had to start with something else, a tool sometimes much stronger: experience.

As all the founders of m19 are ex-Criteo employees, they already had worked on very similar topics: Google Search and Google Shopping campaigns. "When Amazon started its advertising market, we recognized something. There were keywords, products, bids, purchase intent. The structural similarities with Google Shopping were strong," describes Pascal Pons. "I would also like to point out the fact that as we were among the first to use Amazon's first advertising API, it gave us an early-mover advantage in understanding the platform's data structures."

The result was the development of a decision tree-based system. But here lay one of the key differences of m19's solution: it was not a generic ML model pulled off the shelf, but something custom-tweaked for the specificity of Amazon advertising. "The algorithm itself is not rocket science, but we have over-engineered it. And most of the value in machine learning comes not from the sophistication of the mathematics in it, but from the quality of the intuitions embedded in the model's design, the quality of the data it learns from, etc.," continues the computer scientist. "The secret is this: we baked our experience, our knowledge and our intuitions in it. That is what makes it unique."

‍

How m19's AI Actually Works

Even if the original flow of data is scarce and the algorithm a "humble" machine learning one, m19 is acclaimed for its efficiency. A power born from the right usage of something the world of AI has almost erased: statistics. "At m19, we do statistics. The core model ingests 90 days of client data: every keyword impression, every ad click, every attributed sale, every cost per click," explains Pascal Pons. From that historical record, our algorithm learns a single, precise thing: for a given keyword and a given product, if a paid click occurs, what is the probability of a purchase? And if a purchase happens, what will its average value be? No generative magic, no hallucination, no black box. Our model uses the recent past to be as accurate as possible.

But to be efficient, the model must constantly evolve. "The model is retrained every day. Not every week: every day. And it weighs recent events heavily: the last few days matter more than events from two months ago. This is a deliberate choice from us, as it makes the system reactive to seasonality or to market shifts," develops Pascal, who quickly warns: "It won't predict the future! But it will adapt to the present faster than any human analyst could do across a full client portfolio. It's more reactive than predictive, but it's usually faster than a human."

‍

Data Isolation: Why Your Model Never Learns From Other Clients

Could m19 train differently? The answer is yes. Would YOU be happy about that? Pascal Pons is confident the answer would be a firm "no."

"Empirically, the more data you have, the better. We could do that and the extra granularity may enhance the average behavior. But we don't do that. We won't do that. And that's not a technical decision: that's a commercial one," P. Pons explains.

We talk here about one of m19's core architectural decisions: data isolation. Each client's data is kept strictly separate. The model trained on Client A's history never sees Client B's data. The predictions we make for you are derived exclusively from your own behavioral record. "It is true that a larger dataset would, in principle, yield better predictions. But there are two main issues. The first one is the fact that smaller clients would benefit from the knowledge of larger, more established advertisers," continues Pascal Pons. "We don't want to hurt our best customers," he justifies.

The second issue would be raised when two companies compete in the same category. A shared model would implicitly benefit one from the other's data. But the question of who gains more from whom is not answerable in any satisfying way. "We know that it would raise complicated commercial questions. To avoid this issue, we keep the data in silos."

‍

The Cost of Frugality

Whenever AI is involved these days, one of the main questions that comes to mind is the ecological impact of it. And the energy behind it. And the cost behind all of that. Especially when a model is fine-tuned every day. As your success is ours, we had to develop the best technology to keep the advantage against our competition. But because servers are expensive – and so are tokens and energy – we had to think out of the box and develop our own proprietary solution. A system we optimized to the core.

Here, our strength lies in the very nature of our algorithm. "Even when training each client's algorithm every single day on the last 90 days, our ecological footprint is way under some of our competitors. The reason is that our algorithm is a machine learning one. As I said before, it's an overengineered algorithm, designed specifically for its tasks, but it's a tailor made one. Not a generic power-hungry LLM. Execution is extremely frugal, and training takes less than one hour per day," describes Pascal Pons.

In more technical terms: because our machine learning model is purpose-built and runs only on the data it needs, it does not require parallel GPU clusters that power large language models (LLM) or deep learning (DL) systems trained on internet-scale datasets. Hence the exclusive reliance on CPUs, and for a much shorter training window.

‍

m19 AI as a weapon,m19 stance as a shield

By design, running advertising campaigns on Amazon is not a neutral act. The reason is simple: the platform has its own interests. While that's entirely logical, the issue is that their interests do not always align perfectly with those of the advertisers – you. And one of our core missions is to be not an adversary of Amazon, but a professional filter to enhance your TACOS, your revenue while protecting your data.

We have already explained how each client's algorithm is trained exclusively on their own historical data. Another thing our technology does is our unique campaign structures. "We created internal campaign structures that explode the structure so that Amazon's reports give us the most granular useful information."

The key mechanism here is to minimize the number of keywords per campaign unit. Fewer keywords per structure means that the stats that come back from Amazon are more specific. This allows us to isolate performance at the level of a keyword/ASIN combination: "Amazon doesn't give you that by default," adds Pascal Pons. The price to pay for this technology is to prevent you from structuring your own campaigns. A decision we made as our metrics have proven, time and time again, that our internal structure is the most efficient for our clients.

Finally, let's talk about implementation, because as we stated it, technology is not neutral. "We hear from our clients, who are following Amazon announcements and webinars: 'will you integrate this new format? Have you adopted that new bidding feature?' Our clients are demanding and we respect that," acknowledges Pascal Pons. "They want to be up to date with Amazon's ad products."

And m19 answers that call, as Pons explains: "It's where much of the engineering team's time goes. But integration is not automatic, and enthusiasm is not a sufficient reason. We refuse to integrate tools that are not profitable or genuinely interesting for our clients," Pons says. "Before integrating any new functionality, we start asking ourselves, 'does this actually serve advertiser performance, or does it primarily serve Amazon's revenue?' It's only when this is clear that we act for our customers' interests," he concludes.

And when this job is done, our powerful yet very frugal "AI" does the job.

For your benefit, and only yours.