From Systems of Record to Systems of Context — Omri Bruchim & Tomer Ast, monday.com

AI Engineer · 15 min · 186 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:14Um, hey everyone.
00:16Uh, we are super excited to be here.
00:17Thanks for having us.
00:19Uh, today we're going to talk about how we shift monday.com from a system of record uh, into a system of context.
00:27And honestly the title tell the whole story uh in just a single line.
00:32Uh for decades we build uh software that record what's happened um every task, every document, every message um every status update um just put into the record.
00:45What we want to talk today is like take it step further.
00:48Uh we want software that actually understand the connection between them.
00:53So I want to start from a simple question that each one of us ask himself every morning.
00:59Um what should I focus on right now?
01:02Um it sounds almost trivial but uh to be honest with yourself if you ask your agent whether Gemini GPT or even cloud um if you ever typed this question
01:14you probably got list of bullets um not related to each other.
01:19um list of like item dressed up like a a confident paragraph.
01:24Um but but it's not really um um connected to what you're working on.
01:30Actually, I tested last week and uh Cloud asked me to go to the gym.
01:34I don't know if it's a compliment or not, but uh this is what he suggested.
01:38Um and and what really make it so frustrating is like your assistant um have all this data.
01:47It has all the boards, the tasks, the emails, the Slack messages, everything that you ever touched if you connect it.
01:54Um, but it still can't really answer it.
01:57Um, it has all the data.
01:59Um, but it has zero understanding.
02:04Um, so this is the real challenge we are facing.
02:06Um, this is the heart of the entire talk.
02:09The problem was never the missing of data, the retrieval.
02:12The problem is like the missing understanding.
02:15uh those are two totally different things and almost everyone mix between them.
02:21Understanding is the word that we're going to focus the entire the entire talk.
02:25Not context, not memory, not retrieval, understanding.
02:30So um quick about ourselves.
02:32My name is Tormer.
02:33My name is Omri.
02:34This is Tor.
02:35Um uh we both engineering manager at monday.com working on exactly uh the problem that we going to talk about a little a little bit context about where we coming from.
02:45Uh monday.com is a global software company.
02:48Uh we build a work platform uh used by hundreds of thousands of teams.
02:54Um and the part that really mattered for the talk is that Monday is where the work um lives.
03:00uh we help companies to doing the work not just like saving records.
03:04Every project, every task, every decision, every meeting, the notes, the action item, everything logs logged into the system.
03:12And our mission has always been to help the teams to achieve their business outcome.
03:18Whether you are a salesperson, so we want to help you like create an NSDR agent that call to your prospects and help you sell the product.
03:26And if you are like a finance team or marketing we want to help you uh with the research.
03:32So each one of the discipline we want you help you to doing the job and together we have like four big bats on the platform.
03:40We have like Monday sidekick we're going to talk about it mostly today.
03:44Uh Monday vibe if you want to build your own software.
03:47Monday agent if you want to create your own agent into the platform and like if you want some more deterministic flow we have like Monday workflows but today we're gonna focus on psychic
03:57um psychic is your intelligence AI personal assistant that's understand your work think and execute with you like a bird on your shoulder uh he know you he know your business
04:09um it help you with your work on every aspect uh he work the way you do uh with your tones Um um it's keep you totally in control but like
04:20putting all these bullets and uh building all these promise is really hard.
04:24Um so let's talk about why it's hard and there are three points for that.
04:30Um first like picture one assistant sitting on top of absolutely everything.
04:36You have your u your u slack messages.
04:40You have all the notes for your meetings.
04:42um absolutely everything and it's beautiful and overwhelming at the same time because as it stand like there's wall of records everywhere.
04:53Um so what why it's making so hard to connect between them.
04:59There are three reason for that.
05:02One is something that we call the agent gap.
05:05Your agent every agent is really sharp and doing task um when you when he know what to do.
05:11Um but he sometimes lost find them find the problem.
05:14Um if I ask my agent like please help me draft and and and reply to like the escalation from a customer he nailed it.
05:22He know how to do it.
05:24He take the context.
05:25I like doing the work but asking him what should I focus on first he guess because he doesn't understand what is my priority who am I it doesn't matter if you have a memory or something
05:38he still don't know what is the problem the second problem is like um we have the records but we don't have the meaning um a log never said what it mean let's take an analogy from our
05:54Let's say that you have one line of code in your GitHub.
05:58You look at a code, maybe there is a comment on top of it.
06:01Nice.
06:02But you don't really understand why someone wrote this line of code.
06:06Today, no none of us writing code.
06:08But somewhere in the past, someone wrote this code line of code.
06:12And if you really want to know, you can go to get blame and understand uh from the commit log what why why someone do it.
06:20But if you want to go farther, you can go to the PR and maybe read the description.
06:26And if you really want to go hard, you can go to your Monday board and see which PR connected to this item and understand that this line of code came because some customer complain about something.
06:38So this is what we're trying to build.
06:41And the third point that is a bit challenging as well is like um it's really hard to build ahead of time um at runtime um the meaning of things.
06:51Um it's really too late.
06:53Um you simply can't do something like that the moment someone asked the question.
06:59So understanding understanding of the context has to be ahead of time.
07:04Um you need to build it much before someone asks the question.
07:10Um and this is why we have built what we are building.
07:13We are building uh the Monday world model.
07:16This is what we called the Monday world model.
07:18Help you um understand why this matter um how to help you, who you are, when and and what's not to do.
07:28Um it's the context that follow your work.
07:31Um he understand who you are and it's simply not a bigger prompt.
07:38uh it's not a longer uh context window.
07:40It's a totally different um from from what we know until now.
07:45First before like Tor going to talk about how we build it.
07:49What it's not it's not a retrieval problem.
07:51The problem never been getting the data we don't we have all the data we have all the connection to what all the provider that we want to get all the MCPS.
08:01The problem is really to understand how it works.
08:04Understand how each one of these entity connected to each other.
08:08So go ahead.
08:10Thank you.
08:12Hi everybody.
08:13Um so what's the data model?
08:17We collect thousands of data points on the user every item status change their activity log messages and meetings and construct three things the agent can resone over.
08:28The first is how the user's work is structured, their key entities and relationships and connections between them.
08:36What depends on what, how a message in Slack connects to a task, who's blocking whom.
08:42The second is a current snapshot, live signals over those entities, what's overdue, what's critically urgent right now, which co-workers you've been actively working with, and why.
08:54The third is what we can learn about the user over time.
08:57There are decisions and outcomes, work patterns and cadences distilled into a durable profile.
09:05So how do we build that data model?
09:06Um we use two engines running on different time windows and schedules.
09:11A slow engine that runs on a long time window and learns the user and their work and a fast engine that reads what's happening right now and how it affects the user's work.
09:20One knows you and the other one knows your day.
09:24First the slow engine.
09:26It takes as context users activity over weeks and minds it for patterns and the type of persona the user is their routines their work rhythm who they collaborate with their main goals and current projects.
09:42Those patterns get distilled into a durable profile and every time a profile holds it's reinforced.
09:48This engine tries over time to learn who exactly the user is and how they work.
09:56The fast engine is the opposite.
09:58It takes as context a short recent window and recomputes a set of live signals over the user's current state.
10:06What server do, what's suddenly urgent, which co-workers you've been pulled in with.
10:11This engine tries to understand your day and updates frequently.
10:16Uh this split isn't something we invented.
10:19It's present in two totally different fields.
10:22In neuroscience, this split is referred to as complimentary learning systems.
10:26And in data processing architecture, it's referred to as a lambda architecture.
10:31We apply the same concepts to how we construct the agent's data model.
10:37Our brain uses the same split.
10:39Uh every important experience gets captured instantly by the hippocmpus and over time the new cortex distills those into durable lessons.
10:50In data infrastructure, the same there's the same split.
10:54A fast speed layer over a recent real-time window and a slow batch layer over the full history that gets recomputed and the two are merged into a single served view.
11:07Two different fields landed on the same idea and that's what we're trying to apply to our data model.
11:14So how does it all come together?
11:16We collect data from everywhere our users work.
11:19uh Monday, Slack, emails, calendar, and we turn it into data structures, signals, and patterns.
11:27Both engines premputee on top of that offline and ahead of time.
11:32And when a user engages with Sidekick, a thin slice of logic is recomputed for recent activity, and the entire context is served to the agent.
11:42Sidekick can then decide when and how to traverse and retrieve context from the data model itself.
11:48and it's primed to reason on top of it.
11:53And building it this way gives us two behaviors out of the box.
11:57The data model, it's resilient.
11:59Sources are isolated so a bad feed can't break the rest.
12:03And the thin layer of logic that runs at serve time verifies part of the context against live data while the rest falls back to the last verified context.
12:13So it degrades gracefully, but it doesn't fail.
12:18Second, it actually understands the urgency of facts.
12:21It's able to understand when and how it should be proactive and notify the user and when it should stay silent.
12:30And the crucial part is that it compounds.
12:33Every day the data is captured, the layers fill in and the profile sharpens.
12:38And adding a new data source is deliberately cheap and only contributes.
12:43So the surface only grows.
12:45The more it sees, the more it understands.
12:47And the more it understands, the more you can lean on it.
12:51And the data model is unique.
12:52It's unique to how you work.
12:56We're not pretending this solves everything.
12:58Uh the model itself is always trailing the actual live world.
13:04New users have no reliable data to reason from yet, and signals have our own biases built in.
13:11So the hardest part is actually telling the important parts from the noise.
13:16But this is an architectural design that we can enrich and test and improve over time.
13:22Thank you.
13:25[applause]
13:28So back uh back back to the original uh question that we had.
13:32What should I focus on right now?
13:34Remember the the question.
13:35So now Psyche can really answer that.
13:38Uh let's see how.
13:39So first like Thomas said we collect all the data point we called it breadcrumbs for example the board that I working on right now um all the emails from the past uh months
13:50um the transcript for the meeting all the action item that I got from those meeting and we processing them offline um we see a full picture of your calendar this is what help us to understand the pattern so if you have like bi-weekly
14:03with your VPs I understand that tomorrow you have another meetings and even slack messages from the last few days to understand if you say something in the Slack that's relevant that you didn't say in your daily uh standup with the team.
14:17We took all this data and we uh process them either fast or slow.
14:23Um the slow build some kind of like profile on you.
14:26So you can see it on the top like combin is an engineering manager.
14:30We I'm working on two projects psychic and notaker.
14:33I'm from Tel Aviv.
14:34And this is how many hours I have per day and like how I work every day.
14:40And the fast is something that's helped me to understand what is the action item.
14:43I have like three commitment that I promised to other people.
14:46I have to to reply to a VP that sending email and I didn't.
14:50Um so this is only from the past uh uh day or few days window.
14:56This is something that's relevant more for today and this is how psychic can answer those questions.
15:01So to summarize it, the bottleneck was never uh the capability where you can where you take all the data from.
15:08It was the understanding how you connect each one of the dots to each other.
15:12Um the most capable agent in the world whether it's like a cloud and Gemini, it doesn't understand you.
15:18He need to process it beforehand.
15:20This is what we're trying to solve at Psychic.
15:22Um this is the Monday word model.
15:24Uh this is what we're building.
15:26Um if you have any question, uh we out of time.
15:29We are here uh uh near the stage.
15:31Thank you very much for the time.
15:33Uh you can follow us on LinkedIn.
15:34We post things on on the area.
15:37Thank you very much.
15:39[applause]