Freelancer vs Agency for Machine Learning Projects

Where machine learning projects actually spend their time

Choosing between a freelancer and an agency for general software is well covered. Machine learning adds a twist. The modelling, which is what most freelance ML profiles advertise, is a minority of the effort on a production project. Data preparation, integration, deployment and monitoring take the rest.

So the useful question is not 'who is the better data scientist'. It is 'who will do each of those jobs, and who will be there in a year when the model drifts'.

The shares are illustrative and vary by project; the pattern does not.

When a freelancer is the right hire

There are good reasons to hire a freelance ML engineer, and a strong one is often faster and cheaper than any agency for the right job.

• You have in-house developers who will deploy and maintain the model
• The task is narrow and well defined, such as improving an existing model or running an analysis
• The data is already clean and accessible in one place
• You need specialist depth in one area, such as time-series forecasting or computer vision
• The work is exploratory and a single sharp person can answer the question in a few weeks

In those cases an agency adds coordination you do not need. We tell people this on discovery calls, and it costs us nothing we mind losing.

When an agency earns its overhead

An agency makes sense when the project needs several skills at once and nobody in your business can fill the gaps. A typical example is a 60-person distributor wanting demand forecasts inside its ordering system: that needs data engineering to pull sales history from an old ERP, modelling, a service that produces forecasts nightly, changes to the ordering screens and monitoring. One freelancer can do some of that. Very few can do all of it well.

It also matters when continuity is a risk. If the freelancer who built your model takes a full-time job, gets ill or simply stops replying, the model is still running and nobody understands it. An agency such as SpiderHunts documents the work and keeps more than one engineer familiar with it, which is dull and valuable.

The model is rarely what breaks. What breaks is the knowledge of how it was built, when the one person who had it moves on.

The failure modes of each

Freelance ML projects tend to fail in a recognisable way: an impressive notebook, good offline accuracy, and no route to production because nobody planned the integration. The work was good. It just stopped at the part the freelancer was hired for.

Agency ML projects fail differently: too many people, too much process for a small problem, and a junior doing the modelling while a senior sold the work. Ask who will actually do each part, by name. At SpiderHunts the engineer on the discovery call works on the project, and we would expect any agency you consider to say the same.

Questions to ask either one

1. Walk me through a model you put into production and what happened to it six months later
2. How will the model get its data each day, and who builds that?
3. What baseline will you compare against, and what result would make you recommend stopping?
4. How will we know if performance degrades after launch?
5. If you became unavailable tomorrow, what would another engineer need to carry on?
6. Where will the code and trained models live, and who owns them?

A good freelancer answers these as well as a good agency. The answers will tell you far more than portfolios, which in machine learning are often notebooks that never met real users.

Cost, compared fairly

Freelancers usually have lower day rates, and for a contained task they are often the cheaper option overall. The comparison changes when you add the hours your own developers spend integrating the model, the time you spend managing several people, and the risk of rework if the pieces do not fit.

Compare total cost to a working, monitored system in production, not the cost of the modelling phase. Our machine learning service quotes a fixed price for that whole scope, which makes the comparison straightforward. For how we think about pricing models in general, see fixed price versus time and materials.

A combination that often works

Some of the best results come from mixing the two: a specialist freelancer for a hard modelling problem, working inside a structure an agency or your internal team provides for data, deployment and support. If you already have a freelancer you trust, keep them. We are happy to handle the engineering around their model rather than replace them.

Frequently asked questions