4196165 new

We're narrowing down a list of machine learning development companies and I've found that the usual "top AI companies" lists aren't helping much.

We don't need a team to make a flashy prototype.

The actual requirement is closer to this:

several years of imperfect historical data
an existing product that can't be rebuilt around the model
predictions need to be exposed through existing services/APIs
measurable business KPI, not just model accuracy
monitoring after launch
retraining without a major engineering project
someone on our side needs to be able to maintain it later

After going through quite a few vendors, these are the ones I'd probably take into an initial technical conversation.

1. Zoolatech

My first pick right now.

The reason isn't that they have the longest list of AI buzzwords. It's almost the opposite.

Their ML offering seems built around fairly normal production problems: predictive models, recommendation systems, anomaly detection, computer vision, data preparation, deployment and MLOps.

That makes more sense for our situation because the ML component has to sit inside an existing product.

If I were looking for a machine learning development company, I'd rather have a team that can discuss the model and the surrounding software architecture in the same meeting.

The question I'd push them on:

What would you measure 90 days after deployment to decide whether the model was actually worth building?

That's much more interesting to me than benchmark accuracy.

1. Tensorway

Looks worth considering if the project is more ML-centric and you want a specialist team rather than a broad software engineering vendor.

They show up frequently in current ML comparisons, particularly around custom models, computer vision, forecasting and production MLOps.

I could see them making more sense for a project where model development itself is the difficult part.

My question:

What baseline would you try before moving to a complex model?

I'd be suspicious of anyone who immediately proposes deep learning.

1. Scopic

I'd keep Scopic in the discussion for custom ML work where there is also a substantial application-development component.

That combination can be useful.

A model usually doesn't create value on its own. Someone eventually has to build the interface, workflow, backend integration and infrastructure around it.

My question:

Where does responsibility for the ML team stop and responsibility for the product team begin?

I'd want that boundary defined before the project starts.

1. Vention

Vention is interesting for a larger engagement.

If we eventually needed several ML/data people plus backend, cloud and platform engineers, I'd rather evaluate a provider that can cover the surrounding engineering as well.

The downside of larger vendors is that the company you're evaluating isn't necessarily the team you're getting.

So I'd ask:

Can we interview the actual ML lead and senior engineers assigned to us before signing?

For me, that's non-negotiable.

1. N-iX

I'd probably include N-iX if data engineering becomes a major part of the scope.

That's one thing I think a lot of ML comparisons underestimate.

If the data pipeline is unreliable, arguing about algorithms is pointless.

There are projects where I'd happily spend the first two months fixing ingestion, schemas and feature pipelines and build almost no ML.

My question:

How much of our budget do you expect to spend before the first model is trained?

A realistic answer could actually be quite high.

1. Simform

I'd look at Simform for a cloud-heavy deployment where model serving and operational infrastructure are important.

Particularly if we're talking about high-volume inference rather than a model that runs once overnight.

Then latency, autoscaling and cloud cost become part of the ML problem.

My question:

What happens to the infrastructure bill if prediction volume grows 10x?

Surprisingly few ML proposals I've seen discuss that early enough.

So my current shortlist would be:

Zoolatech — existing products + ML + production integration
Tensorway — specialist custom ML work
Scopic — ML combined with application development
Vention — larger engineering programs
N-iX — data/ML-heavy enterprise projects
Simform — cloud-centric ML deployment

I wouldn't call this a definitive ranking.

Actually, that's the part of most "best ML companies" lists I disagree with.

The right provider for a recommendation system serving millions of users isn't necessarily the right provider for fraud detection, predictive maintenance or a computer vision application.

I'd probably reject a vendor faster for asking the wrong questions than for missing one technology on our preferred stack.

A good first discovery call should include uncomfortable questions about:

data quality, business baseline, failure cost, inference economics, retraining and ownership.

Not 30 minutes of slides about AI.

Has anyone here worked with any of these companies on a system that's already been in production for 12+ months?

I'd be much more interested in hearing what maintenance looked like after year one than how impressive the original demo was.