As a human living in 2026, you are drinking from an information firehose.
Slack messages from your colleagues at work; emails from friends or businesses you're interacting with; Whatsapp group chats with your family and neighbors; messages from your kid's school, your doctor, and your bank through specialized apps for each.
The majority of these messages you can skim or just ignore. But a few contain urgent, actionable items for you: a tax form, a notice of a school field trip, an urgent project request from your boss.
It's a huge job to continually triage what's important from this deluge. Can computers help? This is a research question I'll be investigating over the next three months as a researcher-in-residence at Ink & Switch.
Prior art: filtering
Computers enabled this information-overload problem, by making it frictionless to send messages. But relatively little has been done on the receiving side, to help us cope with this flow.
Some prior art includes:
- Spam filtering (e.g. SpamAssassin, Gmail spam filtering): huge success. This is a somewhat easier problem because what constitutes spam is the same for most people.
- Explicit filtering rules (e.g. rules features in many email clients): mostly a failure. It's too much work to set up and then manage these, especially since a given person's preferences will change over time.
- Automated filtering (e.g. Gmail priority inbox, SaneBox): some success, but anecdotally not a must-have or total solution to the problem.
Your personal algorithm
For this project, I want to try a personal algorithm that triages messages for you.
This is inspired by social media platforms. Although there's a lot of negativity about "the algorithm," these systems often do an impressive job at surfacing videos you might want to watch (YouTube, TikTok) or posts you might want to read (Twitter/X).
Importantly, the user doesn't need to manually create rules or manage preferences. Instead it learns from the user's implicit feedback about what they engage with.
The weak point of social media algorithms for end users is that they are only partially designed to serve us. Their larger incentives are to the platform's business interests like maximizing engagement, time on platform, and advertiser attention.
My aim is to make this personal algorithm 100% focused on determining which messages are relevant to your life priorities. Like social media, it should require no manual tuning of rules, only occasional inline feedback of its triage results.
Challenges
I anticipate these challenges:
Training on small datasets. Spam filtering works well because there is a huge global pool of data in Gmail. Each time someone clicks "This is spam" or "This is not spam" they're providing training to a global model that benefits everyone.
This doesn't work for importance filtering, since it's highly personal. One person might love to see an email notifying them of a sale on their favorite brand; for someone else this is an irrelevant distraction. The Gmail priority inbox developers discuss this in their 2010 paper on the subject.
Assume that any given user will have a few hundred or perhaps a few thousand samples in their example dataset, depending on what exactly we use as user feedback signals for training.
The unified inbox problem. There are numerous channels through which you can receive important messages. Even if Gmail priority inbox worked perfectly, that still doesn't cover Slack, Discord, Whatsapp, SMS, social media mentions, specialized apps for medicine/banking/school, or postal mail.
There are products that try to provide unified inboxes for messaging (e.g. Franz) or social media (e.g. Hootsuite). These can be useful in particular cases, but it's an infinitely-deep rabbit hole trying to recreate all the features of each message platform.
Adversarial APIs. Messaging apps and social media often discourage API access of the sort needed to make an alternate interface to their system.
This is partially for business model reasons: Gmail, X, and Facebook Messenger want you to come into their UI where they can serve you ads, collect data about your habits, and engage you with other content.
But this is also because APIs can be used in automated spam and abuse. Discord, for example, doesn't provide any API access to direct messages, and explicitly forbids scraping this data in their terms of service.
Conclusion
This is research and not a product idea, because any of these challenges could make this whole approach non-viable. But I hope to explore far enough to either prove that it's possible in theory, or find and document the dead-end that prevents it from being possible with today's technology.