DOMAIN PERSPECTIVE
Where commerce personalisation is going.
Recommendation is an infrastructure problem wearing a machine-learning hat. Feature freshness, serving latency and cold-start behaviour matter more than the choice of algorithm.
- Latency budgets are measured in tens of milliseconds at request time
- Most catalogue items and most users are cold-start most of the time
- Offline metric gains frequently fail to survive online measurement
WHAT IS ACTUALLY HARD
The problems that decide the architecture.
- 01
Training / serving consistency
One feature definition used by both paths, so the model sees at inference what it saw in training.
- 02
Cold start
Sensible fallbacks for new users and new items, rather than an empty or arbitrary result.
- 03
Honest evaluation
Offline scores treated as a filter, with the real decision made on measured online behaviour.
Work in this domain
Other domains we work in
Have a problem worth solving?
Tell us what you’re trying to build, automate or improve.