Usually, ML interviews consist of case studies that require you to think about what features to use given a use-case. However, interviewers ocassionally also throw open-ended questions like - “How will you increase the functionality of our current product using AI/ML?”. Thus, I had recently given a series of ML interviews for multiple e-commerce platforms and I did some pre-thinking before these interviews to prepare better. Below are a few ways I feel AI can be leveraged to improve the quality of e-commerce apps.

Already happening ✅

- Targetted ads
- Personalized offers
- Dynamic pricing
- Recommendations for increased visibility
- increased engagement

Enhancements 📈

- Rebalance driver fleets according to demand
	- followed by driver allocation
	- incentivize
	- lesser waiting times
	- greater coverage

- augment customer service
	- automate flow with bots
	- provide solutions based on existing knowledge base (customer profiles, previous interactions, existing solutions)
	- greater autonomy to solve problems at care-tech level than engineer level
		- analysis entire order life-cycle to pin-point issue (stuck state, driver reblast)
	- use feedbacks to filter out complaints, suggestions and direct to relevant team
		- summarize pain points
		- provide initial problem suggestions

- augment the app to a more human-like experience
	- "snacks for rainy day"
		- recommend food
			- individual items or set packages
		- provide customized offers (discount on existing cart, discount to add more items)
		- automatically add to cart, review, pay
	- "need ingredients to make biryani"
		- list products, or complete packages
	- promote (users+restos) based on festivals
	- create a mini-google within the app
		- "how much time form X to Y and whats the price"
		- review, book ride
	- spotify-like recommendations
		- "people also like"
		- "top 10 foods in your area"
		- "new releases"

- have a unified cart with multiple restos (like amazon)
	- order pizza from dominos
	- chicken from kfc
	- pay all at once
	- internally calculate respective breakup and pay merchant

- use AI for edge devices
	- "knowledge distillation," or “Domain reduction”, which involves training a smaller model to mimic the behavior of a larger model on a smaller data set
	- quantization to reduce memory footprint
	- hybrid solutions (cloud + edge)
		- with devices running ‘light’ versions of the model for low latency while the cloud processes multiple tokens of the ‘full’ model in parallel and corrects the device answers if needed.

- Robustness