What happens in a project review

A project review is a free 30-minute call about a machine learning project you're scoping or running. We go through it on the Machine Learning Canvas (MLC) — one page for designing an ML system — and look for the gaps that stop projects. You leave knowing where your project is likely to break, and what to check before anyone builds it.

Step 1 · before the call

I draft or mark up your MLC

When you book, share your canvas draft if you've started one, or a few lines about the project if you haven't. If it's for a client, you don't need to name them.

I mark up your draft, or write one from your lines. Every line on it is either what you told me, a guess ending in (my guess), or a question that names who can answer it.

Step 2 · the call

The 30 minutes

0–5 min

I show you the draft and tell you it's probably wrong in places. We discuss the context, what stage the project is at, who owns the business result it's meant to change, etc.

5–20 min

We fill in the rest of the canvas together, and I look for where it breaks.

I ask questions about your data (e.g. testing whether your history has enough examples whose outcome is already known, whether it matches what the model will actually see when it runs).

"We'd have to check" is the most common answer, and it counts as a finding. Expect 1 finding that changes the plan, and 2 questions nobody on the project can answer yet.

20–25 min

I ask what your predictions save and cost. For a yes/no prediction, that's what a correct prediction saves and what a false alarm costs. For a number, like a demand forecast, it's what predicting too high/low costs.

Together, those numbers decide whether acting on the predictions pays off (even when they're wrong). Most teams don't have them yet, so we make a first guess at each one together, from what you know. Every guess gets the name of who can correct it.

Those guesses are where a dollar value for the project starts.

25–30 min

I share my recommendations on next steps.

Then I explain how I work with teams after a review, and what it costs. Working together, I take the costs we guessed through a step-by-step method that turns them into a target for what the system is worth, the performance it has to hit to earn it, and a spec your engineers/agents can build from. We'll discuss whether it's a fit.

Step 3 · after the call

What you get

The same day, I send a short written follow-up you can forward to a partner who wasn't on the call. It holds your MLC as we edited it, the findings in the order you'll want to start them (what needs other people first), and my recommendations on next steps.

I do these reviews for free because reading real projects is the fastest way I have to keep the method honest, and because some of them turn into work we do together.

Book a project review

Pick a time30 minutes, free

Louis Dorard

Creator of the Machine Learning Canvas

The draft MLC
Prediction Task
  • Will this job need a 2nd visit to the same unit within 30 days? Yes or no.
  • About 18% of jobs do.
  • Predicted the evening before, when tomorrow's jobs are assigned.(my guess)
  • The rule today: the oldest units get a senior engineer.(my guess)
Decisions
  • Each evening, the 30% of tomorrow's jobs most likely to need a 2nd visit get a senior engineer and the unit's usual parts.(my guess)
  • Can the dispatcher move a job the list puts on top?→ ask the lead dispatcher
Value Proposition
  • For the dispatcher, who assigns tomorrow's jobs.
  • Fewer jobs need a 2nd visit, because the right engineer and the right parts get there the first time.
  • How does the dispatcher choose who gets a senior engineer today?→ ask the lead dispatcher
Data Collection
  • How far back do job records go, and has the way a 2nd visit gets recorded changed since?→ ask the service manager
Data Sources
  • The job management system, one record per visit.(my guess)
  • Does every visit record which unit it was for?→ ask the service manager
Impact Simulation
  • A 2nd visit costs about $280.(my guess)
  • What does it cost to send a senior engineer and a parts kit to a job that didn't need them?→ ask the finance lead
  • How much would this have to save a year to be worth building?→ ask the operations director
Making Predictions
  • Once a night, for about 95 jobs, before tomorrow's jobs are assigned.(my guess)
Building Models
  • Who on your side would own the model once it's running?→ ask the operations director
Features
  • Unit age, the fault reported, past visits to the same site.(my guess)
Monitoring
  • 2nd visits per 100 jobs, checked monthly.(my guess)
  • Which response times in your service contracts must not slip?→ ask the contracts manager