Resources

Five Questions Before Starting.

A project that cannot answer all five clearly has not yet done the work required to succeed.

1. What does success look like — exactly?

Define minimum acceptable accuracy per class, max false positive rate, max false negative rate, target latency, and evaluation methodology. Write it down. Both parties sign off.

2. What data do you have and what do you need?

How many labelled examples per class? Who will annotate? How long will it take? Does the data cover your deployment conditions? Budget 3× the naive annotation estimate.

3. What is the gap between a demo and production?

A working baseline is the start of the project, not the end. Closing the gap to production quality requires multiple iteration cycles. Does your timeline and budget reflect this?

4. What is the realistic timeline?

Use the realistic estimates, not the naive ones. Build in buffer for data issues — they always occur. Define phase deliverables so progress is visible before problems become crises.

5. What happens six months after deployment?

Who monitors performance? What triggers retraining? How will deployment failures feed back into data? How will updated models be validated? Plan this before deployment, not after.

← Return to ArticleBrowse All Resources →