Aampe conducts controlled, parallelized experiments with tagged messages to learn user preferences.

Aampe

Aampe’s agentic infrastructure enables product and marketing teams to build strong customer relationships by delivering experiences that are always personalized. Once deployed, Aampe’s agents continuously learn user preferences and optimize the delivery of messages and in-app experiences. 

For every user, Aampe assigns an agent that uses machine learning and human guidance to continuously learn about its client – the user – and decide what to deliver, when to deliver, and most importantly, whether or not to deliver in the first place. 

Built by a team of empathetic and experienced data scientists and engineers, Aampe serves marketing, growth, and product teams at consumer and prosumer technology companies. 

Aampe has successfully helped household brands across Europe, Asia, and North America to amp up their personalization game. 


Accenture recently published a fascinating and powerful research report on "Cutting through the noise in consumer experience". 

They reported that "74% of consumers abandoned their shopping baskets in the last three months simply because they felt bombarded by content, overwhelmed by choice and frustrated by the amount of effort they need to put in to making decisions." 

At Aampe we know that THE fundamental problem for every consumer business today is effectively and empathetically managing their customers' attention. We live in an attention-scarce economy and the businesses that take care of their customer's 'attention account' will be the ones that will sustainably access their customer's financial account. 

With Aampe's Content and Message Catalogue capabilities, your CRM and product teams can finally engage every single user in their own context, with the content that reduces their stress, and improves their buying journey.

Reach out today for a demo and to find out more!

Aampe is hosting a mini-series of live masterclasses on Recommender Systems starting next week! 

The first session will have DJ Rich and Schaun Wheeler talk about what really matters when implementing a Recommender System. Arpit Choudhury will orchestrate 🥁 

If you search for information on Recommender Systems, everything has been SEO-optimized to talk about general principles – collaborative filtering, content-based recommendations, etc – and delves into specifics like matrix factorization. None of this helps a business get a Recommender System up and running.

This session has been designed specifically for product and growth practitioners to learn what really matters when implementing a Recommender System. 

When?
August 14, 11:30 am ET (3:30 pm GMT)

Register now: https://lu.ma/7i9sgzdz

We've been working on making improvements to our recommender system. 

(Aampe lets you enable as many recommender systems as you like - our agents learn individual user preferences for each system - but if you don't already have your own recommender system, you can just use ours.)

I won't go into all of the details, but we've implemented something similar to Amazon (see link to their paper in the comments below), which involves a random permutation of all of the user-item associations. So imagine a big spreadsheet where first column is a user id and the second column is an item id and every time a user interacts with a new item, you add a row to the sheet. We needed to randomly shuffle the item id column of that sheet. 

However, we're not just talking about a siple random shuffle. We need two things to be true at the same time:

1. The number of users associated with an given item should be roughly what it was before we randomly shuffled anything.
2. No user should ever be associated with the same item twice.

All of this is easy to do in Python, because it's easy to work iteratively in Python. We can shuffle all the item ids, check to see which users have duplicated ids in their new set, then shuffle again to replace those duplicates, and so on until we have the result we want. The problem is that we get well over a million unique user-item associations in a single week, and we need to ingest much more than one week to really explore that data enough to make robust recommendations. The Python route is simpler to code but it comes at a relatively large I/O cost from moving all that data out of and then back into databases.

So I spent some time this last week figuring out a database-only solution. A quick very just uses the LAG or LEAD operator, with an ORDER BY RAND() statement in the window specification. However, LAG and LEAD lack a wrap-around functionality, so one value will always be left null. If that's unacceptable, then you have to ROW NUMBER your whole dataset, then create one or more alternate row numbers with ORDER BY RAND() in the window spec, and then self-join the table on different row number columns. To solve the duplicates issue, we self-joined on five different random columns, aggregated all of those together into a single array per user, and then selected the number of item ids we needed to meet that user's quota.

At any rate, we now have the random permutation we need for our recommender system (and, in fact, we're looking into ways to expand this method to other parts of our agentic learning). And now we don't have the burden of moving lots of data back and forth between two storage systems.

Have you ever thought about value-based personas instead of demographics?

Understanding the psychology of values unlocks fascinating insights into why we connect so deeply with certain people and brands. When individuals share similar values, they often experience a natural, effortless bond, leading to meaningful personal and professional relationships.

Companies can harness this powerful dynamic by weaving their core values into every aspect of their brand. This attracts customers who resonate with these values and fosters loyalty and trust. 

At Aampe, we do exactly that. During the onboarding process, we identify the company values and why customers are using your app. What engages them? Once we know that, we build messages that speak to every user. 

Join me on Thursday for our webinar about value-based personas and how we identify them!

"It means that we get to label and generate content, and then the machine learning technology behind the scenes 'learns' how to distribute that content to engage and convert.

Agent Al is the distribution part of messaging. It acts like an agent, performing tasks and making decisions based on data and predefined rules. This helps automate and optimize the process of determining whom to send messages to, when to send them, and how often to send them.

Semantic machine learning is the content part. It uses machine learning to understand the meaning and context of data (semantics), building upon that predictive modelling about what content to send and whether to send it at all.

So looking forward to learning and building up the solution and foundation, which will not only help reduce friction which will finally help properly distribute different types of content (and reduce the manpower needed)."



Great summary, Vernesa. We couldn't have said it better ourselves! 😊

You and the team have also been an absolute pleasure to work with. Human creativity is an instrumental part of the process, and you guys never fail to bring it!

👀 Want a sneak peek at what we're working on at Aampe? 👀

Look no further than Schaun Wheeler's latest post, below! 👇

COM:AAMPE

Aampe

Filings Not Available