Meeting Title: Client Hub n8n Deep Dive Date: 2025-10-23 Meeting participants: Hannah Wang, Samuel Roberts
WEBVTT
1 00:00:33.000 ⇒ 00:00:34.160 Hannah Wang: Hello!
2 00:00:34.320 ⇒ 00:00:35.690 Samuel Roberts: Hey, how are you?
3 00:00:36.340 ⇒ 00:00:44.570 Hannah Wang: Good. I’m at a cafe again. I feel like whenever I call you, I’m outside, so… so if it’s loud, then…
4 00:00:44.790 ⇒ 00:00:46.119 Samuel Roberts: No worries, I can’t hear anything.
5 00:00:46.810 ⇒ 00:00:47.669 Samuel Roberts: Hopefully you can hear me.
6 00:00:47.670 ⇒ 00:00:49.709 Hannah Wang: AirPods. Yes, I can.
7 00:00:50.030 ⇒ 00:00:50.970 Samuel Roberts: Okay, great.
8 00:00:50.970 ⇒ 00:00:57.330 Hannah Wang: Okay, so… I don’t know if we’ve ever done a case study together.
9 00:00:57.400 ⇒ 00:00:58.480 Samuel Roberts: No, I don’t think so.
10 00:00:58.730 ⇒ 00:01:03.819 Hannah Wang: Okay, well, this is, like, slightly different than a case study. It’s gonna be more of, like, a technical…
11 00:01:04.230 ⇒ 00:01:05.140 Hannah Wang: I guess…
12 00:01:05.470 ⇒ 00:01:15.440 Hannah Wang: like, blog… we’re gonna try to make a technical blog post slash whitepaper out of… out of whatever we discussed today. I think that’s what Bhutan mentioned, so…
13 00:01:15.440 ⇒ 00:01:30.809 Hannah Wang: I know that we’re going to talk about the client hub, and just, like, I guess, deep diving into it, particularly with the technology, the tools that we use, like NAN and, like, the agent flows, and all the technical stuff. So, I…
14 00:01:30.840 ⇒ 00:01:36.369 Hannah Wang: for… this is my first time doing, like, a technical project, quote-unquote, interview. I usually do.
15 00:01:36.370 ⇒ 00:01:36.760 Samuel Roberts: Okay.
16 00:01:36.760 ⇒ 00:01:45.040 Hannah Wang: these for case studies, I feel like they should be pretty similar. Essentially, I’m just gonna ask you a bunch of questions, and even if
17 00:01:45.150 ⇒ 00:01:56.570 Hannah Wang: you have to repeat yourself, or even if it’s obvious, like, still explain it, because I take the transcript of this meeting, I feed it to AI, and then I output, like, the content for
18 00:01:56.570 ⇒ 00:02:13.169 Hannah Wang: the case study or the blog post. So, yeah, I’ll just guide you through, a bunch of questions, and then you could also share your screen if one… if and when applicable, if there’s, like, a good visual you can show me. And… yeah, let’s dive right in. So…
19 00:02:13.170 ⇒ 00:02:18.099 Samuel Roberts: Okay, fair warning, let me know if I’m getting, like, too technical, or I’m not deemed technical enough.
20 00:02:18.250 ⇒ 00:02:22.870 Samuel Roberts: Because I’m not 100% sure how deep to go different times for this.
21 00:02:23.050 ⇒ 00:02:24.629 Hannah Wang: No, go all in.
22 00:02:24.800 ⇒ 00:02:27.629 Hannah Wang: Even if I don’t understand, that’s okay, because…
23 00:02:28.250 ⇒ 00:02:32.689 Hannah Wang: Yeah, the audience is for technical people, and I’m not super technical, so…
24 00:02:32.690 ⇒ 00:02:41.989 Samuel Roberts: Okay, I’ll go as much as I can then. The other side of it is, like, I’ve been working with them a little bit, but, the NAN stuff, but I didn’t necessarily implement a lot of this stuff to begin with.
25 00:02:41.990 ⇒ 00:02:43.859 Hannah Wang: Okay. So, no worries.
26 00:02:44.160 ⇒ 00:02:50.960 Samuel Roberts: I can give you some stuff, but if you want more, then talking to Casey Mustafa at some point for this same piece might be helpful.
27 00:02:51.350 ⇒ 00:02:55.920 Hannah Wang: Fair. Okay, I’ll probably have them review it as well, and then give.
28 00:02:55.920 ⇒ 00:02:56.310 Samuel Roberts: Great.
29 00:02:56.310 ⇒ 00:02:57.530 Hannah Wang: feedback, so…
30 00:02:57.680 ⇒ 00:03:07.350 Hannah Wang: Okay, so, number one, I guess, is just the context and background. So, like, why did we even build this client hub in the first place?
31 00:03:07.350 ⇒ 00:03:13.970 Samuel Roberts: Sure, okay, yeah, so this is, this also, predates me at Brainforge a little bit, but, you know, the basic…
32 00:03:15.430 ⇒ 00:03:28.469 Samuel Roberts: The basic purpose is that there’s lots of different meetings, there’s lots of information kind of floating around from those meetings, and we want to be able to have a good kind of central location that is
33 00:03:28.850 ⇒ 00:03:34.689 Samuel Roberts: you know, keeping track of the transcripts from those meetings. It also includes, Slack.
34 00:03:35.560 ⇒ 00:03:53.529 Samuel Roberts: channels, so, you know, you can go into these client hubs and ask what’s currently going on, and get, you know, more recent information. You can talk about previous things, but basically, it’s kind of centralizing as much of the kind of client context that we have for, you know, it was very helpful for me getting up and running, for example.
35 00:03:53.900 ⇒ 00:04:10.039 Samuel Roberts: on a few different projects. But the idea is effectively just trying to bring as much of the information into one place as possible, using the transcripts from the meetings, Slack information. Those are kind of the big two big ones right now, but we’re looking to add more pieces of information, like documents and things.
36 00:04:10.590 ⇒ 00:04:12.169 Hannah Wang: So I guess the pain point…
37 00:04:12.510 ⇒ 00:04:15.169 Hannah Wang: I guess this is for, like, our team internally,
38 00:04:15.170 ⇒ 00:04:15.670 Samuel Roberts: Yes.
39 00:04:15.670 ⇒ 00:04:26.329 Hannah Wang: But, like, I guess the pain point is that every… like, knowledge is dispersed everywhere, and we just wanted to get it consolidated into one place, right? Like, that’s the pain point that people were…
40 00:04:26.910 ⇒ 00:04:28.560 Hannah Wang: Like, I guess, mentioning…
41 00:04:28.900 ⇒ 00:04:41.879 Samuel Roberts: I think, yeah, I think there’s that, and I think this is, you know, one way to go about that, especially for things that are a little more ephemeral, like meetings or Slack messages, where, you know, you can have a kind of centralized Notion, you know, wiki-type thing.
42 00:04:41.880 ⇒ 00:04:55.019 Samuel Roberts: But, obviously that requires keeping it up to date. Right. That requires people knowing where to go, and updating things, knowing… the new people knowing where to go for that. And so the thing about the Client Hub is that it pulls in lots of information that’s…
43 00:04:55.470 ⇒ 00:04:57.769 Samuel Roberts: Probably would get lost from that if it weren’t…
44 00:04:58.200 ⇒ 00:04:58.910 Hannah Wang: Hmm.
45 00:04:58.910 ⇒ 00:05:00.700 Samuel Roberts: for this, I think.
46 00:05:02.260 ⇒ 00:05:11.319 Hannah Wang: Okay, cool. The second part is, I guess, the technical architecture of everything, and just, like, the
47 00:05:11.490 ⇒ 00:05:21.859 Hannah Wang: the infra and the framework and stuff. So, what tools did we use to build, this client hub? I’m assuming NAN, what else?
48 00:05:21.860 ⇒ 00:05:27.919 Samuel Roberts: And it ends a big one, yeah, it’s doing a lot of the, the flows for not only,
49 00:05:28.620 ⇒ 00:05:33.800 Samuel Roberts: Using the client hub, but also the creation of each new one, which is a big part of the process right now.
50 00:05:34.150 ⇒ 00:05:37.050 Samuel Roberts: But essentially, yeah, we have our Zoom meetings.
51 00:05:37.510 ⇒ 00:05:43.900 Samuel Roberts: Get recorded, transcribed, that all gets processed into Excuse me, Superbase.
52 00:05:44.280 ⇒ 00:05:48.020 Samuel Roberts: And embeddings are generated from that, so it’s more easy to search.
53 00:05:48.920 ⇒ 00:06:05.840 Samuel Roberts: Same thing with the Slack messages, those are getting ingested into Superbase, and embeddings are created for that. And then N8N is where, when you make a request, either from a Slack bot right now, or from the, the Forge, the platform.
54 00:06:06.300 ⇒ 00:06:23.370 Samuel Roberts: you can query all of that information. It’s a little, complex, as you can imagine, for N8N. You know, there’s lots of different flows and different things, but the essential core logic is that there’s an AI agent that knows about these tools, specifically the
55 00:06:24.150 ⇒ 00:06:35.740 Samuel Roberts: Superbase Zoom transcripts and the Superbase Slack messages, and when you ask a question, it uses that, its knowledge of the client, there’s some, kind of…
56 00:06:36.070 ⇒ 00:06:53.949 Samuel Roberts: stuff hard-coded into the system prompt about the clients, and about the people working on, and things like that, but then a lot of it comes from the actual meeting transcripts in Slack, and then it kind of does a quick search on that, pulls that back into context, passes that through a little summarizer, and gives you the information you’re looking for.
57 00:06:55.510 ⇒ 00:06:56.960 Hannah Wang: So right now, it’s…
58 00:06:57.380 ⇒ 00:07:13.489 Hannah Wang: it pulls from the Slack mess… like, any Slack threads or Slack messages, and also Zoom. I… wait, sorry, you mentioned the Forge? Yes. What… how is that a part of the client hubs? Because I’m… the client hubs are in Slack, right? Like, we… at… Yes. Client Hub. Yes.
59 00:07:13.490 ⇒ 00:07:18.959 Samuel Roberts: It can do that, but it also, when you’re… when you’re looking at the… when you go to the client pages on the Forge…
60 00:07:18.960 ⇒ 00:07:20.120 Hannah Wang: Oh, okay.
61 00:07:20.120 ⇒ 00:07:22.780 Samuel Roberts: That is hooked up to the same kind of data source, so…
62 00:07:22.780 ⇒ 00:07:23.230 Hannah Wang: I see.
63 00:07:23.230 ⇒ 00:07:38.850 Samuel Roberts: It’s effectively the same thing. There’s a few slight differences in the way it works in N8N, but it’s basically hitting the same kind of endpoint that’s doing the same searches, doing the same summary, and then coming back with the information.
64 00:07:39.210 ⇒ 00:07:46.690 Hannah Wang: Okay, and just so I know, NAN is, like, workflow… I guess, work… a workflow builder, essentially, kind of?
65 00:07:46.690 ⇒ 00:07:59.230 Samuel Roberts: Yeah, yeah, it’s, you know, kind of a node-based, you know, you put in pieces that are like, when this webhook happens, the output flows to this field, or this thing that edits the field, that flows into this thing that goes to the AI agent.
66 00:07:59.230 ⇒ 00:08:12.639 Samuel Roberts: And then that AI agent node is connected to a bunch of different tools, so it knows to call that, and then when that one’s done, it sends that output further to something that post-processes it, and then sends it back to Slack or to the webhook or whatever.
67 00:08:12.680 ⇒ 00:08:15.860 Samuel Roberts: So, yeah, it’s a kind of node-based workflow editor.
68 00:08:16.380 ⇒ 00:08:18.670 Hannah Wang: And then what is… what is Supabase?
69 00:08:19.030 ⇒ 00:08:21.080 Samuel Roberts: Superbase is the database layer.
70 00:08:21.080 ⇒ 00:08:21.840 Hannah Wang: I see, okay.
71 00:08:21.840 ⇒ 00:08:28.000 Samuel Roberts: It’s Postgres, PostgreSQL, so, Superbase is just the provider of that database, and it has some.
72 00:08:28.000 ⇒ 00:08:28.750 Hannah Wang: Yes. Okay.
73 00:08:28.750 ⇒ 00:08:36.460 Samuel Roberts: hooks kind of built into N8N, because you can… you can do it all, but, Supabase specifically connects very well with N8N.
74 00:08:36.460 ⇒ 00:08:37.260 Hannah Wang: Got it.
75 00:08:37.260 ⇒ 00:08:38.109 Samuel Roberts: Oh, yeah.
76 00:08:38.110 ⇒ 00:08:43.770 Hannah Wang: Okay, and I guess, what makes this, like, approach to…
77 00:08:43.890 ⇒ 00:08:47.890 Hannah Wang: like, a client hub type of thing? Like, what… what makes it unique? Like.
78 00:08:48.360 ⇒ 00:08:52.989 Hannah Wang: What about the architecture, or, like, what about the approach that we used?
79 00:08:53.210 ⇒ 00:08:58.180 Hannah Wang: Makes it unique, and, like, how does it compare to maybe other…
80 00:08:58.540 ⇒ 00:09:11.549 Hannah Wang: I guess, hubs that people try to build. I know you mentioned something about, like, oh yeah, you can manually update, like, a doc or a Notion page. Obviously, it’s ephemeral, and yeah, I guess, like, what about this approach?
81 00:09:11.720 ⇒ 00:09:18.050 Hannah Wang: Both from, like, a idea perspective, but also, like, a technical perspective, like, makes it unique.
82 00:09:18.460 ⇒ 00:09:32.350 Samuel Roberts: Yeah, so, from the idea perspective, I think it’s really nice to have something that’s keeping track of all these meetings and messages, in a way that is queryable. You know, you can build, kind of, RAG pipelines
83 00:09:32.380 ⇒ 00:09:40.560 Samuel Roberts: You know, retrieval augmented generation stuff, which can do all kinds of things with documents, but the idea of then plugging in our actual
84 00:09:40.730 ⇒ 00:09:44.179 Samuel Roberts: Transcripts, conversations of meetings,
85 00:09:44.260 ⇒ 00:09:53.399 Samuel Roberts: the Slack messages, the threads there, I think lets you catch up really quickly to things that you might not be able to otherwise. It frees up time.
86 00:09:53.470 ⇒ 00:10:04.609 Samuel Roberts: that you might have to pull someone and say, like, hey, what happened in this meeting? You know, I see the… maybe I… maybe you can see the agenda, or you can see the output of the meeting, but you might miss some context that’s there. And so the ability to have…
87 00:10:04.940 ⇒ 00:10:15.169 Samuel Roberts: this kind of query ability into those things. In addition to some of the stuff we have on the platform, like, you can watch the meeting as well, you can see the.
88 00:10:15.170 ⇒ 00:10:15.560 Hannah Wang: Right.
89 00:10:15.660 ⇒ 00:10:25.670 Samuel Roberts: But then you can also kind of go a little bit beyond that and say, like, okay, when did this come up in other meetings? When did this come up in Slack? What’s still outstanding? That’s… that is something that I don’t think is really…
90 00:10:25.990 ⇒ 00:10:33.050 Samuel Roberts: doable without having to get a lot of people wrangled to kind of find that information on the fly.
91 00:10:34.330 ⇒ 00:10:38.479 Samuel Roberts: Technically, it’s essentially, you know, a RAG system kind of mapped onto…
92 00:10:38.710 ⇒ 00:10:51.510 Samuel Roberts: transcripts, that are being added to constantly. So, you know, I wouldn’t say it’s anything kind of crazy there, but I think the real, you know, it’s not… we could have built the right system on top of, say, like, Notion Docs, or on top of other things.
93 00:10:51.510 ⇒ 00:10:51.930 Hannah Wang: Hmm.
94 00:10:51.930 ⇒ 00:11:01.499 Samuel Roberts: But again, I think the fact that it’s meetings that are getting added, there’s a whole other layer of ingesting these meetings every time there’s a meeting, and Slack messages periodically, and things like that.
95 00:11:01.610 ⇒ 00:11:05.620 Samuel Roberts: That keeps it up to date, keeps it… current.
96 00:11:05.820 ⇒ 00:11:13.549 Samuel Roberts: you know, has things that are the most topical currently for, you know, what’s going on, or what happened over the last X number of days, or things like that.
97 00:11:13.690 ⇒ 00:11:15.699 Samuel Roberts: That I think is… is a really…
98 00:11:16.280 ⇒ 00:11:24.899 Samuel Roberts: it’s more beneficial than just, like, okay, here’s a bunch of documents, search over it and tell me what’s going on. It’s… here’s a bunch of other contexts that we might lose otherwise.
99 00:11:25.070 ⇒ 00:11:25.740 Samuel Roberts: So…
100 00:11:26.760 ⇒ 00:11:27.540 Hannah Wang: Gotcha.
101 00:11:27.640 ⇒ 00:11:35.280 Hannah Wang: And I know you were, like, when I asked you about, the tools that were used, you kind of went into the flow, but.
102 00:11:35.480 ⇒ 00:11:36.429 Samuel Roberts: I guess…
103 00:11:36.520 ⇒ 00:11:43.340 Hannah Wang: maybe you can share your screen of the actual, like, flow and NAN, or wherever it’s built, and just, like, walk me through… Yeah.
104 00:11:43.340 ⇒ 00:11:44.560 Samuel Roberts: Technically.
105 00:11:44.720 ⇒ 00:11:47.629 Hannah Wang: Yeah, go super nitty-gritty.
106 00:11:47.630 ⇒ 00:11:51.929 Samuel Roberts: Okay, yeah, so there’s, let me, let me share a…
107 00:11:52.890 ⇒ 00:11:58.710 Samuel Roberts: One of the client hubs. Which one is this?
108 00:11:59.300 ⇒ 00:12:00.919 Samuel Roberts: This one’s a little better to look at.
109 00:12:01.240 ⇒ 00:12:04.480 Samuel Roberts: I think this is the window, so… .
110 00:12:05.540 ⇒ 00:12:06.210 Hannah Wang: Yep.
111 00:12:06.820 ⇒ 00:12:07.910 Samuel Roberts: So…
112 00:12:08.390 ⇒ 00:12:24.849 Samuel Roberts: Yeah, so this is N8N, and you can kind of see the nodes and the different flows, but effectively what’s happening here is… I think this is the right one to share, but I might bounce around from a few of them, to be honest, but… Sure. Effectively, what you can see here, the big things are messages from Slack and this webhook.
113 00:12:24.980 ⇒ 00:12:33.929 Samuel Roberts: So these are kind of the inputs, or the triggers, I guess, is the real term. And so you can see there’s some processing to figure out
114 00:12:34.200 ⇒ 00:12:39.109 Samuel Roberts: you know, the thread and stuff, and you can do… it does a bunch of lookup and things. This is all kind of standard…
115 00:12:39.310 ⇒ 00:12:47.579 Samuel Roberts: you know, N8in is not necessarily the only thing that lets you do this, but what is nice is that we can plug these… both of these into this AI agent here.
116 00:12:48.020 ⇒ 00:12:52.450 Samuel Roberts: This chat message one is for, like, testing over here, so it is helpful for…
117 00:12:53.190 ⇒ 00:13:01.559 Samuel Roberts: development, but it’s not really one of the utilized ones, but these other two flow right into this AI agent, which, if we open up, you can see…
118 00:13:01.690 ⇒ 00:13:05.880 Samuel Roberts: We got the system instructions here, you were a client hub for Brainforge AI, I believe…
119 00:13:06.390 ⇒ 00:13:08.680 Samuel Roberts: This one,
120 00:13:08.980 ⇒ 00:13:14.020 Samuel Roberts: Yeah, it’s a bunch of information here. I’m trying to see if it says which one this is. This is… this might be the…
121 00:13:14.250 ⇒ 00:13:18.109 Samuel Roberts: One that we copy every time, to start.
122 00:13:18.110 ⇒ 00:13:18.760 Hannah Wang: Excuse me.
123 00:13:18.760 ⇒ 00:13:26.059 Samuel Roberts: But the big thing is that this agent has, obviously, this information. We can give it the prompt of everything that’s coming in.
124 00:13:26.230 ⇒ 00:13:29.460 Samuel Roberts: And then it has these two tools here.
125 00:13:29.650 ⇒ 00:13:35.279 Samuel Roberts: You can also see it’s got the chat models that are hooked up to it, so what’s… one thing that’s nice about N8in is that it lets you
126 00:13:35.420 ⇒ 00:13:55.160 Samuel Roberts: pretty easily swap those in and out, so not just for testing, but also for, you know, in the… as we move, like, you know, you can test this with OpenAI and then move to our Azure instances, for example, and so you can see here we have all these hooked up to our Azure OpenAI instances, for doing the embeddings and for doing the chat. But the, these two here are actually what query
127 00:13:55.170 ⇒ 00:13:56.709 Samuel Roberts: Superbase for us.
128 00:13:57.120 ⇒ 00:14:02.260 Samuel Roberts: And so… These are just… these are actually blocks that come…
129 00:14:02.500 ⇒ 00:14:12.840 Samuel Roberts: built in. So this is where I was saying that Superbase is just Postgres, but it’s nice that it has this layer that connects well with N8N. So, you know, it’s…
130 00:14:13.310 ⇒ 00:14:27.960 Samuel Roberts: convenient in that sense, and it’s a nice way to use these tools together. Kind of the ecosystem of all these tools together that work well together is very nice, because it’s not impossible to do otherwise, it’s very possible to do, but it would be a little more…
131 00:14:28.570 ⇒ 00:14:30.189 Samuel Roberts: Bespoke, or per…
132 00:14:30.190 ⇒ 00:14:30.790 Hannah Wang: Right.
133 00:14:30.940 ⇒ 00:14:40.269 Samuel Roberts: And so then this just queries the Zoom Superbase and the Slack Superbase, or it has the ability to, these are all tools that it can call, just like any other.
134 00:14:40.380 ⇒ 00:14:42.689 Samuel Roberts: AI agent,
135 00:14:43.450 ⇒ 00:14:48.979 Samuel Roberts: And then, I think they’re even listed in here. Oops, I don’t have to do that. Oh, I can’t search, sorry. Let me open this up bigger.
136 00:14:51.500 ⇒ 00:14:52.540 Samuel Roberts: Yeah, see?
137 00:14:52.690 ⇒ 00:15:00.560 Samuel Roberts: Oh, no, okay, it’s not searching very well, but you can see a lot of the stuff that’s, like, hard-coded in here. These are some things that are… one thing that is…
138 00:15:00.660 ⇒ 00:15:20.369 Samuel Roberts: good about this is the ability to, like, quickly iterate, and it is really good for that. I personally have found that we’re probably hitting a little bit of the limit of what we can do here, or want to do here, because, you know, this is not the most complex one that we have. You know, you can see there’s some other ones, this might not be the best example, but,
139 00:15:21.290 ⇒ 00:15:27.599 Samuel Roberts: I can show you some other end flows that are even more complex, multi-agent flows,
140 00:15:28.200 ⇒ 00:15:33.750 Samuel Roberts: Which may be worth talking about a little bit. Yeah. So this one’s pretty simple. Let me pull up the other one.
141 00:15:34.160 ⇒ 00:15:37.769 Samuel Roberts: That, I was thinking about before we started talking.
142 00:15:38.130 ⇒ 00:15:41.150 Samuel Roberts: So that is… this guy… oops.
143 00:15:42.200 ⇒ 00:15:45.449 Hannah Wang: Real quick, what is an agent? Like.
144 00:15:45.450 ⇒ 00:15:51.009 Samuel Roberts: Oh, yeah, no. Yeah. Sure, sure, sure. The agent is, is, is basically the, the AI
145 00:15:51.190 ⇒ 00:16:04.340 Samuel Roberts: LLM call. So, let me, let me jump back real quick. Sorry, yeah, I kind of glossed over that. So, this AI agent node is a little different than the other nodes, if you notice. So, like, most nodes have, like, one input.
146 00:16:04.440 ⇒ 00:16:10.400 Samuel Roberts: Maybe a couple more inputs, depending on what it’s doing here. Maybe a couple outputs, like you can see this code one.
147 00:16:10.560 ⇒ 00:16:23.109 Samuel Roberts: is, like, a success and an error, but the agent has, like, an input, an output, and then a bunch of other things you can hook up to it. And so, this agent node is effectively, just like when you want to use ChatGPT, there’s, like, a
148 00:16:23.210 ⇒ 00:16:29.299 Samuel Roberts: your prompt, will you ask it? Behind the scenes, there’s a system message that kind of tells it what it’s…
149 00:16:29.670 ⇒ 00:16:30.680 Samuel Roberts: doing.
150 00:16:30.970 ⇒ 00:16:33.529 Samuel Roberts: Gives it some context of…
151 00:16:34.090 ⇒ 00:16:37.119 Samuel Roberts: Before, just this message is asked, you know, what that is.
152 00:16:37.370 ⇒ 00:16:42.909 Samuel Roberts: So this kind of explains that it’s a client hub, and what it’s going to have access to, and what its purpose is.
153 00:16:43.740 ⇒ 00:17:02.829 Samuel Roberts: And then, these tools, you have, the chat model is actually what’s getting hit for the LLM, so this all gets kind of processed together and passed to that model. The tools are, you know, what it has the ability to make a request to, so it’ll, you know, you might say a prompt like, what happened this week on Slack?
154 00:17:03.450 ⇒ 00:17:20.380 Samuel Roberts: this is what then looks at that, figures out it needs to make that call to one of these tools. That tool gets called, sends back the data, and then it processes it all here, too. So, it’s doing a few things, that might be different pieces, potentially, if you weren’t doing it in N8N.
155 00:17:20.540 ⇒ 00:17:23.610 Samuel Roberts: But, it’s kind of a slightly more complex node.
156 00:17:23.800 ⇒ 00:17:24.560 Samuel Roberts: than just the…
157 00:17:24.569 ⇒ 00:17:28.429 Hannah Wang: I see. Standard ones. But effectively… So, multi… Sorry, go ahead.
158 00:17:29.160 ⇒ 00:17:36.240 Hannah Wang: Oh, finish your statement. Oh, I was gonna say, like, it’s just, it’s a little different than kind of just the standard input-output ones. Yep.
159 00:17:36.460 ⇒ 00:17:45.659 Samuel Roberts: But, besides that, it’s effectively, like, input does a few other things, sends an output, it’s just a little more complex under the hood using the LLMs, which is
160 00:17:46.890 ⇒ 00:17:52.039 Samuel Roberts: More… or, sorry, less deterministic than some of these other ones that are just, like, code, or just…
161 00:17:52.490 ⇒ 00:17:55.470 Samuel Roberts: process a workflow, or read a message, or something, so…
162 00:17:55.470 ⇒ 00:17:56.070 Hannah Wang: Right.
163 00:17:56.370 ⇒ 00:18:07.709 Hannah Wang: So I’m assuming multi-agent means there’s multiple of these, so we can kind of go into the other… Oh, yeah, yeah, exactly, exactly. So let me… so this is a slightly more complex one. That’s crazy, yeah.
164 00:18:07.710 ⇒ 00:18:21.799 Samuel Roberts: Yeah, and this is… this is where you kind of see, like, why it’s good and bad, I find, and ADN is really nice for being able to, like, see everything, and you can kind of get an understanding that you might not have if all this was just in code.
165 00:18:21.930 ⇒ 00:18:25.619 Samuel Roberts: But at the same time, it becomes, like, visually complex in a way that…
166 00:18:25.620 ⇒ 00:18:26.320 Hannah Wang: Hmm.
167 00:18:26.320 ⇒ 00:18:34.300 Samuel Roberts: you know, you can restructure code, and… I mean, you can do different things in N8N2 to, like, break these into sub-workflows and things like that, but,
168 00:18:34.530 ⇒ 00:18:41.280 Samuel Roberts: Yeah, so effectively, this one is a multi-agent, and you can see, like, depending on what kind of thing is coming in here.
169 00:18:41.400 ⇒ 00:18:44.329 Samuel Roberts: We can process it through different agents.
170 00:18:44.470 ⇒ 00:18:48.089 Samuel Roberts: This one, I’m trying to think…
171 00:18:48.360 ⇒ 00:18:50.630 Samuel Roberts: How best to explain this, but the real…
172 00:18:50.990 ⇒ 00:18:53.839 Samuel Roberts: Multi-agent part here is this guy.
173 00:18:54.020 ⇒ 00:18:56.949 Samuel Roberts: So, you can see here.
174 00:18:57.130 ⇒ 00:18:59.759 Samuel Roberts: There’s still an AI agent node, which is the same…
175 00:19:00.110 ⇒ 00:19:13.030 Samuel Roberts: kind of node we had before, but this now goes to a bunch of other nodes that we’ve defined, that are basically the same thing, in terms of their AI agents.
176 00:19:13.280 ⇒ 00:19:18.489 Samuel Roberts: but it… This agent is, like, a routing agent, effectively.
177 00:19:19.060 ⇒ 00:19:28.299 Samuel Roberts: And so, for this one, it would say, like, you know, some prompts would come in, some input, it would… it pulls all the prompts, it then…
178 00:19:29.240 ⇒ 00:19:34.639 Samuel Roberts: processes through all of these different agents as it thinks it needs them. So,
179 00:19:35.570 ⇒ 00:19:39.840 Samuel Roberts: I’m trying to think how deep to get into this, like, for the actual application versus just the…
180 00:19:39.840 ⇒ 00:19:41.099 Hannah Wang: the technical side of it.
181 00:19:41.270 ⇒ 00:19:41.820 Hannah Wang: Go deep.
182 00:19:41.820 ⇒ 00:19:46.449 Samuel Roberts: Okay, okay. So, like, for this one, this is for a client that,
183 00:19:46.870 ⇒ 00:20:00.880 Samuel Roberts: you know, is doing, like, a version 1 of some copy, and so, you know, there’s a narrative agent here that is kind of the overall flow of something. This is the copywriter that actually is generating a copy, the QA compliance that is actually doing the…
184 00:20:01.150 ⇒ 00:20:07.459 Samuel Roberts: Double-checking and making sure that the sources are accurate and things like that, and that we’re not just hallucinating stuff.
185 00:20:07.550 ⇒ 00:20:13.150 Samuel Roberts: And so, like, this agent is responsible for then calling all of these other agents as needed.
186 00:20:13.220 ⇒ 00:20:30.309 Samuel Roberts: And so this is where it gets, you know, a little less, like I said, deterministic, a little more… it figures out, okay, let’s… let’s call the narrative agent, because that’s, you know… In here, we may have… I’m not sure… oh, we have the prompt, sorry. This one we reworked so we could edit the prompts, and another tool.
187 00:20:31.730 ⇒ 00:20:38.710 Samuel Roberts: But… let’s go to this orchestrator prompt, it’s not going to be here, so I can’t actually show you the prompts, because they’re in another tool, I can pull that up if we need to, but…
188 00:20:38.860 ⇒ 00:20:44.860 Samuel Roberts: That was another nice thing about NAN, is that it had some of these other nodes that let us pull prompts from another
189 00:20:45.210 ⇒ 00:20:49.810 Samuel Roberts: tool that let us version control those.
190 00:20:50.240 ⇒ 00:20:56.359 Samuel Roberts: Which was nice, because that is one thing I think NNN does have the ability to do some version control at…
191 00:20:57.160 ⇒ 00:21:07.399 Samuel Roberts: another level, and it then can also be self-hosted, there’s a whole bunch of other ways you can use it, and so different ways we’re using it have different restrictions and stuff, but…
192 00:21:07.400 ⇒ 00:21:07.850 Hannah Wang: Yeah.
193 00:21:07.850 ⇒ 00:21:08.520 Samuel Roberts: Mom.
194 00:21:08.940 ⇒ 00:21:12.999 Samuel Roberts: This one, yeah, we were able to pull these prompts in, feed them all into where they need to go.
195 00:21:13.060 ⇒ 00:21:29.259 Samuel Roberts: And then, yeah, this basically passes it into another agent that we call the steel man that’s just kind of checking to make sure that we rate things, like how well is it doing what we asked it to. We generate some overall feedback that the client asked for that they kind of use as their rubric.
196 00:21:29.310 ⇒ 00:21:37.459 Samuel Roberts: So you can see here we have 1, 2, 3, 4, 5, 6, 7… or 6, 7, 8… 8 different agents, all doing different things.
197 00:21:38.360 ⇒ 00:21:39.669 Samuel Roberts: You can see how this can get…
198 00:21:39.990 ⇒ 00:21:44.030 Samuel Roberts: powerful, but also very complex to manage. Right.
199 00:21:44.550 ⇒ 00:21:51.950 Samuel Roberts: And so, I think this is, like, a benefit and a drawback a little bit into different… depending on what you’re trying to do, you know what I mean? Yeah.
200 00:21:52.490 ⇒ 00:21:56.380 Samuel Roberts: So I think for some of these things, like, we’re hitting a little bit of the limit of what is…
201 00:21:56.970 ⇒ 00:22:04.139 Samuel Roberts: not possible with NAN, because I think a lot of things are possible, I just don’t know if it’s necessarily the best idea, the best tool for the job for everything like this, because…
202 00:22:04.140 ⇒ 00:22:04.520 Hannah Wang: Right.
203 00:22:04.520 ⇒ 00:22:12.319 Samuel Roberts: You can see, like, we have… we’re calling all these other agents for other reasons. It’s… it’s… it’s… this one got very complicated, and I…
204 00:22:12.860 ⇒ 00:22:13.480 Hannah Wang: Yeah.
205 00:22:13.480 ⇒ 00:22:18.379 Samuel Roberts: Had a hard time even keeping up with what was happening at different times as we were making changes, so…
206 00:22:19.510 ⇒ 00:22:22.690 Samuel Roberts: Yeah, so that’s NAN, kind of…
207 00:22:23.500 ⇒ 00:22:25.670 Samuel Roberts: high-end, low level, I guess, in different ways.
208 00:22:28.050 ⇒ 00:22:31.839 Samuel Roberts: I don’t know if there’s more… like, tell me what else is good to get into, I guess, here.
209 00:22:31.840 ⇒ 00:22:50.490 Hannah Wang: Sure, yeah, I’ll just kind of dig into my other questions, and we’ll see where that takes us, so… Perfect, perfect. I know you mentioned, like, yeah, particularly for… I’m assuming this is Interlude, for Interlude, like, this agent, it was kind of difficult, because you have to manage, like, all these agents, and yeah, clearly it’s a complex workflow.
210 00:22:50.490 ⇒ 00:22:51.880 Hannah Wang: Is this, like…
211 00:22:52.240 ⇒ 00:22:59.099 Hannah Wang: I guess my next question is, like, what’s the hardest technical challenge you faced while building these agents? And, like, if Interlude is…
212 00:22:59.260 ⇒ 00:23:07.600 Hannah Wang: one of them, like, feel free to dive into it. If not, like, yeah, what was the hardest technical challenge building client hubs?
213 00:23:07.810 ⇒ 00:23:11.510 Samuel Roberts: The hardest technical challenge…
214 00:23:12.050 ⇒ 00:23:15.820 Samuel Roberts: So, the client hubs, I wasn’t really around when that first… the first, kind of.
215 00:23:16.190 ⇒ 00:23:16.590 Hannah Wang: Right.
216 00:23:16.590 ⇒ 00:23:35.549 Samuel Roberts: version of this was implemented. Like I said, this is kind of a version of a fairly standard RAG flow, so, I don’t think there’s a ton here. What I think got really complex for… for Interlude specifically was, and using NATN, this is maybe one of the…
217 00:23:35.630 ⇒ 00:23:41.620 Samuel Roberts: Drawbacks is, testing and evaluation.
218 00:23:42.490 ⇒ 00:23:43.960 Samuel Roberts: become a little…
219 00:23:45.110 ⇒ 00:24:01.340 Samuel Roberts: hard to do. So as you can see here, like, we’re looking at the editor right now. You can also see all the executions of this, which is nice. It’s great to get in here and really, like, debug specific flows. And so, like, you know, we can click on this one, and it’ll pull up,
220 00:24:03.190 ⇒ 00:24:20.269 Samuel Roberts: If it’s… yeah, okay. So it pulls up, like, what got called, like, this message in Slack, triggered, and if you… on some of the flows, it’s easier to see than others, but let me zoom in here. You can see, like, this is the one that got called, you can watch the flow all the way through. This one hit the if statement, and it looks like it got…
221 00:24:21.610 ⇒ 00:24:31.889 Samuel Roberts: Next, because I think this is catching all the messages from Slack, not just the ones that are kind of for… So, like, this one just, like, is like, okay, that’s not what we’re looking for.
222 00:24:32.200 ⇒ 00:24:36.490 Samuel Roberts: But, you know, evaluations is a whole other thing here where you have to set up
223 00:24:36.600 ⇒ 00:24:41.090 Samuel Roberts: Datasets, and what we… we were actually using some other tools to do that for this.
224 00:24:41.190 ⇒ 00:24:44.620 Samuel Roberts: Because… and it didn’t seem like the ideal…
225 00:24:45.030 ⇒ 00:24:48.269 Samuel Roberts: Place to do some of these, kind of, end-to-end tests of the tool.
226 00:24:48.270 ⇒ 00:24:49.589 Hannah Wang: Oh, okay.
227 00:24:49.590 ⇒ 00:24:57.139 Samuel Roberts: So we were looking at some other things, I’m trying to remember what we were working at for Interlude at this point. I think it was called Brain Trust was one of them. Yep.
228 00:24:57.140 ⇒ 00:24:58.200 Hannah Wang: Yep,
229 00:24:59.210 ⇒ 00:25:15.719 Samuel Roberts: Braintrust was one, that we got into. The other one that we started using, which was not for the evaluations, but was actually for this prompt management, was LangFuse. All of these tools, like, have different… slightly different offerings that all overlap in different ways.
230 00:25:15.860 ⇒ 00:25:16.830 Hannah Wang: Right. So…
231 00:25:16.830 ⇒ 00:25:20.490 Samuel Roberts: And they’re all evolving, because everything’s moving so quickly in the space, but…
232 00:25:20.490 ⇒ 00:25:21.080 Hannah Wang: You know.
233 00:25:21.080 ⇒ 00:25:25.269 Samuel Roberts: Brain Trust is really good, I don’t think I’m even logged in here, so I don’t know.
234 00:25:25.620 ⇒ 00:25:29.770 Samuel Roberts: Yeah, let me see if I can show you real quick what was in here, but…
235 00:25:30.040 ⇒ 00:25:36.409 Samuel Roberts: This was nice, because we could actually give it kind of, like, the golden… data set.
236 00:25:36.790 ⇒ 00:25:44.559 Samuel Roberts: process some things. It was a little complex to make work with N8N. You know, so the idea with N8N is that you have these nodes that are…
237 00:25:44.770 ⇒ 00:25:45.560 Samuel Roberts: kind of…
238 00:25:46.060 ⇒ 00:25:59.240 Samuel Roberts: built by people at those companies, there are some community nodes, there’s other things you can do. But if, you know, there’s also just, like, make a webhook request, or receive a webhook request. So you can really plug them into a lot of different things.
239 00:25:59.350 ⇒ 00:26:04.090 Samuel Roberts: it just becomes a little more complicated than it is for other things.
240 00:26:04.470 ⇒ 00:26:06.680 Samuel Roberts: So, I’m just trying to see if I can show you, like…
241 00:26:06.930 ⇒ 00:26:14.070 Samuel Roberts: you know, this one, this data was coming out, this data was coming out, forget exactly how we passed this in, but I think somewhere in this…
242 00:26:14.400 ⇒ 00:26:21.239 Samuel Roberts: you know, this is the other thing, is like, where in NNN is that getting called? I have to kind of… Yeah. And so,
243 00:26:21.520 ⇒ 00:26:24.640 Samuel Roberts: I don’t remember specifically, because it’s been a minute since I was in here, but…
244 00:26:24.920 ⇒ 00:26:27.770 Samuel Roberts: Somewhere in here, this output is getting fed.
245 00:26:28.210 ⇒ 00:26:33.190 Samuel Roberts: to, you know, Brain Trust, but it’s not happening in NAN, it’s happening in Brain Trust, and it’s.
246 00:26:33.190 ⇒ 00:26:33.770 Hannah Wang: Yeah.
247 00:26:33.770 ⇒ 00:26:38.159 Samuel Roberts: This is where it gets a little… all the tools kind of play nicely in some respects, but not in every respect.
248 00:26:38.270 ⇒ 00:26:47.500 Samuel Roberts: Same thing with LangFuse. LangFuse has some of the similar things to Brain Trust, but what we really liked about it was the prompt version control, and the fact that it.
249 00:26:47.500 ⇒ 00:26:47.830 Hannah Wang: had an.
250 00:26:47.830 ⇒ 00:26:52.469 Samuel Roberts: node built in that would just let us pull the right prompt and not have to do any custom
251 00:26:53.090 ⇒ 00:26:55.399 Samuel Roberts: API requests or things like that.
252 00:26:55.710 ⇒ 00:27:00.260 Samuel Roberts: You know, all things that are doable in different ways, But…
253 00:27:01.020 ⇒ 00:27:07.890 Samuel Roberts: And it then sometimes makes it easy, like, for this one, for example, very easy to just make this node for BrainTrust.
254 00:27:08.190 ⇒ 00:27:14.839 Samuel Roberts: much more complicated. Yeah. And partly because it has its own evaluations tool that we didn’t really like, partly because it’s…
255 00:27:15.040 ⇒ 00:27:17.330 Samuel Roberts: It’s limited in some other ways, but…
256 00:27:18.520 ⇒ 00:27:23.620 Samuel Roberts: That, I think, is, like, where we’re hitting a little bit of the limit here, and where some of the things got a little more complex.
257 00:27:23.930 ⇒ 00:27:24.640 Hannah Wang: Right.
258 00:27:25.330 ⇒ 00:27:29.319 Samuel Roberts: But… Yeah, I kinda lost the thread there a little bit, I’m sorry, but…
259 00:27:29.320 ⇒ 00:27:35.290 Hannah Wang: No, you’re good. Yeah, so basically, you kind of had to, like, piecemeal different tools together in order to, like.
260 00:27:35.620 ⇒ 00:27:38.919 Hannah Wang: I guess, test and, like, evaluate whether the agent was.
261 00:27:39.120 ⇒ 00:27:39.949 Samuel Roberts: Exactly, yeah.
262 00:27:39.950 ⇒ 00:27:41.220 Hannah Wang: End to end, basically.
263 00:27:41.220 ⇒ 00:27:51.479 Samuel Roberts: Right, right. And like I said, for certain things, you know, and it is good for piecing those together, but for certain other things, it’s… it’s doable, but it’s more complex than it…
264 00:27:52.530 ⇒ 00:27:56.969 Samuel Roberts: Maybe could be or should be, for different other… other tools.
265 00:27:57.650 ⇒ 00:28:00.469 Hannah Wang: So, if you had to, like, rebuild this agent.
266 00:28:01.190 ⇒ 00:28:03.970 Hannah Wang: Would you do anything differently, or is this, like.
267 00:28:04.480 ⇒ 00:28:07.759 Hannah Wang: Is this, like, the best option we have, I guess?
268 00:28:07.760 ⇒ 00:28:13.080 Samuel Roberts: Yeah, so, we’re actually in the middle of moving some of this Client Hub stuff into code. Oh, okay.
269 00:28:13.660 ⇒ 00:28:23.720 Samuel Roberts: So, one of the things about this is… I don’t know if you can see all the different ones I have here, like, this is for one client, this is ABC, this is default, this is hip, this is interlude.
270 00:28:24.170 ⇒ 00:28:26.830 Samuel Roberts: When we get new clients, these are all kind of…
271 00:28:28.450 ⇒ 00:28:33.579 Samuel Roberts: tailored to those clients, even though it’s all the exact same logic. And so…
272 00:28:33.880 ⇒ 00:28:44.490 Samuel Roberts: I’m sure there’s a way in N8N to generalize these, but at this point, this logic is so set, we kind of want to just, like, harden it into code that is, a little more…
273 00:28:46.440 ⇒ 00:29:02.870 Samuel Roberts: what’s the word I want to hear? Like, this is obviously within our control, but this is running on NAN servers, we want to kind of have that… like, now that… NAN is great for prototyping, it’s good for even building, but I think codes, in my mind, still wins for kind of, like, final version.
274 00:29:02.870 ⇒ 00:29:03.340 Hannah Wang: Hmm.
275 00:29:03.340 ⇒ 00:29:12.570 Samuel Roberts: Especially as, like, things get more complex. Like, if I had known the deck agent was gonna get this complex when we first started, we might have gone a slightly different route, but it was not this complex the first time, you know what I mean?
276 00:29:12.570 ⇒ 00:29:13.080 Hannah Wang: Yep.
277 00:29:13.080 ⇒ 00:29:16.469 Samuel Roberts: Like, V1 of this was much smaller.
278 00:29:16.470 ⇒ 00:29:17.180 Hannah Wang: Yeah.
279 00:29:17.180 ⇒ 00:29:21.549 Samuel Roberts: The client hub stuff, though, we’re definitely hitting a point. I think I mentioned a little bit about the,
280 00:29:22.700 ⇒ 00:29:24.670 Samuel Roberts: The creation of these client hubs?
281 00:29:25.000 ⇒ 00:29:36.029 Samuel Roberts: is, a bit tedious, because you have to duplicate this, you have to go into all these things and update them for the right client, and things like that. You know, things that…
282 00:29:36.670 ⇒ 00:29:45.110 Samuel Roberts: I would like to now make this more generic and put into some code, so we’re looking at that now. We actually are in the middle of kind of ticketing out some of those things to…
283 00:29:45.630 ⇒ 00:29:47.609 Samuel Roberts: Use some other tools that are…
284 00:29:47.910 ⇒ 00:29:53.620 Samuel Roberts: you know, gonna run right alongside the forge, so things should hopefully be even a little faster, too, because that’s the other thing.
285 00:29:53.620 ⇒ 00:29:53.980 Hannah Wang: Yes.
286 00:29:54.250 ⇒ 00:29:57.820 Samuel Roberts: Some of these flows take… let me see if I can see real quick…
287 00:29:58.360 ⇒ 00:30:00.530 Samuel Roberts: Oh, this is not the one to look at.
288 00:30:03.340 ⇒ 00:30:06.359 Samuel Roberts: Oh, hold on. Is it loading? What’s going on here?
289 00:30:07.010 ⇒ 00:30:09.790 Samuel Roberts: Did my whole thing freeze? Yep, reload, there we go.
290 00:30:10.630 ⇒ 00:30:14.930 Samuel Roberts: Some of these executions can take quite a while.
291 00:30:15.100 ⇒ 00:30:24.969 Samuel Roberts: These ones are not crazy, but you can imagine some of our other N8N flows that are even more complex, like, maybe the decation’s a good one to look at.
292 00:30:26.670 ⇒ 00:30:33.050 Samuel Roberts: I don’t know if anyone’s called this in a while with a real command here, but some of them get up to, like, 20 seconds,
293 00:30:33.050 ⇒ 00:30:33.610 Hannah Wang: Oh, wow.
294 00:30:33.610 ⇒ 00:30:40.130 Samuel Roberts: Minutes, even. Like, it’s… it’s crazy, and there’s not a lot of insight into what’s going on while that’s happening.
295 00:30:40.370 ⇒ 00:30:41.740 Hannah Wang: Yeah. Which…
296 00:30:42.200 ⇒ 00:30:43.140 Samuel Roberts: is…
297 00:30:43.320 ⇒ 00:30:48.980 Samuel Roberts: fine for quick flows, but for longer ones, it’s a little bit like, did it break? Is it just stalling? Is it hitting a…
298 00:30:49.090 ⇒ 00:30:50.909 Samuel Roberts: A bad endpoint, you know, there’s a lot of…
299 00:30:50.910 ⇒ 00:30:51.470 Hannah Wang: Yay.
300 00:30:51.650 ⇒ 00:30:53.210 Samuel Roberts: uncertainty in there.
301 00:30:53.340 ⇒ 00:31:00.920 Samuel Roberts: While it’s happening, and so we’re looking to move some of those over. But yeah, the client hub specifically, because of the way they all get.
302 00:31:01.050 ⇒ 00:31:03.240 Samuel Roberts: Duplicated and recreated, which…
303 00:31:03.400 ⇒ 00:31:10.029 Samuel Roberts: you know, worked initially, and you know, I’m not sure I would have done it differently to start.
304 00:31:10.310 ⇒ 00:31:16.489 Samuel Roberts: But as they become more and more used, and more and more, Kind of.
305 00:31:18.580 ⇒ 00:31:22.200 Samuel Roberts: re… re… you know, re… repeating ourselves a lot.
306 00:31:23.010 ⇒ 00:31:26.830 Samuel Roberts: I think it’s finally time to, like, move this into code and generalize them a little bit.
307 00:31:27.180 ⇒ 00:31:27.900 Hannah Wang: Hmm.
308 00:31:28.650 ⇒ 00:31:47.009 Hannah Wang: Gotcha. Yeah. Okay, so I know we kind of went on a tangent and talked about other agents, but going back to the Client Hub, I guess going… moving into, like, the impact and results of the Client Hub, so I know you mentioned that even for you, it was, like, helpful when you were onboarding to certain clients,
309 00:31:47.360 ⇒ 00:31:54.670 Hannah Wang: I guess, in general, like, what changed after we created this client hub? Like, was… was it, like.
310 00:31:54.680 ⇒ 00:32:08.710 Hannah Wang: I’m sure, like, when people use it, like, I’m sure they… it’s faster to get onboarded to clients, it’s… you can get more clarity, more up-to-date, like, information, real time, because it’s pulling from the latest
311 00:32:08.900 ⇒ 00:32:17.050 Hannah Wang: Zoom meetings and Slack messages, like, what… I guess, like, what impacts did the client hubs have in general, for our team?
312 00:32:17.660 ⇒ 00:32:36.099 Samuel Roberts: That is, I mean, everything you said there is kind of what I… again, I was not here pre-Client hubs. Right. So, I don’t necessarily know what it was like before, but I can imagine, because I’ve been at places that didn’t have things like this, and getting up to speed, and catching up on a meeting you might have missed, or, you know…
313 00:32:37.550 ⇒ 00:32:45.410 Samuel Roberts: I mean, you weren’t involved in that something got mentioned that you needed to know. All that kind of stuff is lost a lot of other places.
314 00:32:45.410 ⇒ 00:32:45.740 Hannah Wang: Right.
315 00:32:45.740 ⇒ 00:32:57.130 Samuel Roberts: Or, again, you need to then get with someone who was there, or check the output of the meeting, which doesn’t have all the context. There’s a lot of things, I think, that, we… I’ve just been kind of making use of them.
316 00:32:59.010 ⇒ 00:33:03.410 Samuel Roberts: And, and not realizing, like, what I would have done previously. Yeah.
317 00:33:03.870 ⇒ 00:33:06.299 Samuel Roberts: But yeah, I don’t have a great sense of, like.
318 00:33:06.720 ⇒ 00:33:12.480 Samuel Roberts: What specifically was like, oh, we need to do this for, you know, this is falling through the cracks, or whatever.
319 00:33:12.480 ⇒ 00:33:13.990 Hannah Wang: That’s okay.
320 00:33:13.990 ⇒ 00:33:24.939 Samuel Roberts: But I definitely can see how useful they’ve been. For me, I definitely, like, especially, like, onboarding, onto new, like, clients, and onboarding onto the company in general, getting up to speed on things.
321 00:33:24.960 ⇒ 00:33:34.799 Samuel Roberts: I found that very helpful in catching up on things, or going back, you know, knowing that everything that’s being said in the meeting is going to be available there definitely helps me…
322 00:33:34.930 ⇒ 00:33:41.930 Samuel Roberts: you know, I still like taking notes for my own personal need, sometimes, but…
323 00:33:42.460 ⇒ 00:33:50.050 Samuel Roberts: I also know that, like, okay, I can go double-check this. I don’t know if you ever have this experience where, you know, you’re watching YouTube videos, and you go back, and you go back, and then you’re
324 00:33:50.430 ⇒ 00:33:53.830 Samuel Roberts: on, like, TV, and it doesn’t have any, like, go-back feature, and you’re like, oh.
325 00:33:53.830 ⇒ 00:33:54.600 Hannah Wang: Yeah.
326 00:33:54.600 ⇒ 00:33:59.710 Samuel Roberts: really got used to that without realizing it. Yeah. The client hubs have definitely done that for me.
327 00:34:00.500 ⇒ 00:34:08.920 Samuel Roberts: Where I’m just like, I appreciate the ability to jump back and go back to a meeting that was yesterday that I remember, but don’t remember the exact wording of something, for example.
328 00:34:08.920 ⇒ 00:34:09.480 Hannah Wang: Right.
329 00:34:09.480 ⇒ 00:34:10.650 Samuel Roberts: Things like that, yeah.
330 00:34:11.199 ⇒ 00:34:21.069 Hannah Wang: this… like, I don’t know how I went to school anymore. Like, you can’t rewind, like, the pre-lecture recordings, like, pre-Zoom, like, in middle school, high school, like, how did we…
331 00:34:21.229 ⇒ 00:34:23.899 Hannah Wang: learn. We had to, like, take notes real time, and then.
332 00:34:23.900 ⇒ 00:34:24.730 Samuel Roberts: Yeah.
333 00:34:24.730 ⇒ 00:34:27.170 Hannah Wang: If we missed it, oh well, like, you have to…
334 00:34:27.179 ⇒ 00:34:38.579 Samuel Roberts: It’s crazy, it’s crazy to think about, like, how much was just, like, lost there, because I didn’t jot something down, or what I missed, or the, you know, inability to, like, you know, make my notes match up with what was actually said after the fact.
335 00:34:38.609 ⇒ 00:34:40.649 Hannah Wang: Especially if you have bad handwriting like me, it’s a…
336 00:34:42.270 ⇒ 00:35:00.119 Hannah Wang: Totally. Okay, the last thing is just, like, reflections and takeaways for you. I think we’re gonna write this blog post as if you, like, and you’re gonna be the author of it, on our website and stuff, so yeah, just getting your reflections and takeaways, like, yeah, what did you learn about
337 00:35:00.270 ⇒ 00:35:10.750 Hannah Wang: like, NAN in general, like, while building client hubs and even just, like, digging into all the workflows and stuff? Like, yeah, is there anything that you learned about
338 00:35:11.030 ⇒ 00:35:13.169 Hannah Wang: AI agents and workflows and NAN.
339 00:35:13.170 ⇒ 00:35:20.940 Samuel Roberts: Yeah, yeah, so any of that, I had, like, some experience before, coming from, like, a home automation kind of background.
340 00:35:21.290 ⇒ 00:35:33.320 Samuel Roberts: Where some node builders like this are really helpful for certain things, and it then, with the AI stuff, definitely kind of blew up that way, and I got a little bit of taste of it, but seeing, like, what we’ve been able to do with it definitely
341 00:35:34.890 ⇒ 00:35:40.769 Samuel Roberts: I have a deep appreciation for it, for, like, what it can do, and where its limitations are now.
342 00:35:41.110 ⇒ 00:35:41.790 Hannah Wang: Hmm.
343 00:35:41.810 ⇒ 00:35:44.240 Samuel Roberts: I definitely also think I have a…
344 00:35:44.410 ⇒ 00:36:00.730 Samuel Roberts: biased towards code, because that’s what I’ve been more comfortable with. Just historically, that’s what I’ve done. And so, trying to think, like, is this a good use case for N8N? Is it… when is it going to stop being a good use case? You know, prototyping? It’s great, you can change things so quickly, you can rerun things, we don’t have to worry about
345 00:36:00.860 ⇒ 00:36:06.169 Samuel Roberts: Lots of, like, the infrastructure and, like, where code’s running, and it works on my machine kind of things.
346 00:36:06.590 ⇒ 00:36:19.059 Samuel Roberts: But, you know, planning things out ahead of time, knowing, like, for example, that deck agent for Interlude, as it got more complex, might have been something that I should have realized sooner, was like…
347 00:36:19.260 ⇒ 00:36:23.550 Samuel Roberts: harder to do an N8N, longer term.
348 00:36:24.880 ⇒ 00:36:35.140 Samuel Roberts: you know, I think the client hubs, when they were created, everything made sense to do in N8N, but as we’ve learned more about duplicating them for new clients, or
349 00:36:36.140 ⇒ 00:36:39.189 Samuel Roberts: You know, how slow things can run, how fast things can run.
350 00:36:39.750 ⇒ 00:36:49.290 Samuel Roberts: I think I… I have a better sense of now, great, this is good for, like… like, certain things will still probably stay in N8N, but I also think there’s something to be said for…
351 00:36:49.700 ⇒ 00:36:53.079 Samuel Roberts: Building smaller workflows that get triggered by other things.
352 00:36:53.520 ⇒ 00:36:54.270 Hannah Wang: Hmm.
353 00:36:54.270 ⇒ 00:37:00.590 Samuel Roberts: whether that’s our code, on the platform, whether that’s other N8in flows. I think, you know.
354 00:37:01.020 ⇒ 00:37:11.200 Samuel Roberts: well-defined tasks are great for it. Ai usage for prototyping, I think, is also really good. I think even production, depending on how complex it is, is really good.
355 00:37:11.800 ⇒ 00:37:14.529 Samuel Roberts: But… Excuse me, as complexity gets…
356 00:37:15.350 ⇒ 00:37:21.460 Samuel Roberts: bigger, I think N&M becomes a little bit of a behemoth to manage sometimes.
357 00:37:22.190 ⇒ 00:37:36.369 Samuel Roberts: But that also might just be my bias towards something visual versus something code. You know, I also think I appreciate people that can maybe do any of them better than I can, versus, like, I might code better than… you know, there’s, like, there’s a push and a pull there where
358 00:37:36.370 ⇒ 00:37:43.230 Samuel Roberts: I wouldn’t knock any in for any of the things that I’ve talked about, I think it’s just, you gotta pick the right tool for the right job, like anything else.
359 00:37:43.230 ⇒ 00:37:43.900 Hannah Wang: Yeah.
360 00:37:44.080 ⇒ 00:37:46.120 Samuel Roberts: And that is very dependent on, like.
361 00:37:46.480 ⇒ 00:38:03.449 Samuel Roberts: what people are comfortable with. You know, the right tool for one person might be the wrong tool for someone else, depending on their aptitude or their need. You know, basically what I’m saying is, like, I like N8in for certain things, and I think that might overlap with other people in certain ways and not others. Yeah.
362 00:38:03.830 ⇒ 00:38:08.919 Samuel Roberts: But… Yeah, I mean, that’s kind of all I had to say there. I don’t know if that’s.
363 00:38:08.920 ⇒ 00:38:10.690 Hannah Wang: Yes, yeah, that’s great.
364 00:38:10.930 ⇒ 00:38:19.540 Hannah Wang: Yeah, I mean, that… that goes for, like, any… anything, even outside of, like, technical stuff. It’s like, oh, some things work for some people, some things don’t, and…
365 00:38:19.540 ⇒ 00:38:19.860 Samuel Roberts: that.
366 00:38:19.860 ⇒ 00:38:22.610 Hannah Wang: How you learn and how you process things.
367 00:38:22.990 ⇒ 00:38:24.780 Samuel Roberts: And I think… I think this is really, like.
368 00:38:25.100 ⇒ 00:38:33.280 Samuel Roberts: solidified that for me even more, because it was a tool that I was less experienced with, and kind of probably would have written off as more of a prototyping thing in general.
369 00:38:33.280 ⇒ 00:38:34.440 Hannah Wang: I don’t know.
370 00:38:34.440 ⇒ 00:38:44.170 Samuel Roberts: And now I still think it’s very good for prototyping, and I can also see it good for production use cases, but I also now know… I have a better sense of, like, where that line might be for as we’re building new things.
371 00:38:44.440 ⇒ 00:38:47.640 Samuel Roberts: That I didn’t have before, so, yeah.
372 00:38:48.530 ⇒ 00:38:51.180 Hannah Wang: Cool. Well, yeah, I think…
373 00:38:51.270 ⇒ 00:39:05.189 Hannah Wang: Those were all the questions. I asked AI to help me come up with questions to ask you, so we’ll see, we’ll see what the output is of this blog post, or, like, technical white paper type of thing, but yeah, I appreciate your time. I…
374 00:39:05.190 ⇒ 00:39:13.070 Hannah Wang: we’ll probably… me or Anne will probably mock this up, get the content, and then we’ll have you review it before we kinda…
375 00:39:13.410 ⇒ 00:39:15.520 Hannah Wang: Go into high fidelity.
376 00:39:15.520 ⇒ 00:39:15.870 Samuel Roberts: Okay.
377 00:39:15.870 ⇒ 00:39:18.980 Hannah Wang: But yeah, we’ll probably ask you to look over the contents, and…
378 00:39:18.980 ⇒ 00:39:19.430 Samuel Roberts: Great.
379 00:39:19.430 ⇒ 00:39:23.439 Hannah Wang: Let us know if it’s, like, completely off, because we’re going to use AI, obviously.
380 00:39:23.440 ⇒ 00:39:24.400 Samuel Roberts: Yeah, no.
381 00:39:24.400 ⇒ 00:39:25.740 Hannah Wang: content, so…
382 00:39:25.740 ⇒ 00:39:29.789 Samuel Roberts: I definitely, yeah, no, I get that. I will definitely take a look at it, make sure it makes sense.
383 00:39:29.790 ⇒ 00:39:34.929 Hannah Wang: Okay, cool. But yeah, thanks for your time, and I will…
384 00:39:34.930 ⇒ 00:39:35.540 Samuel Roberts: Yeah, thank you.
385 00:39:35.540 ⇒ 00:39:37.060 Hannah Wang: IT on Slack. Alright.
386 00:39:37.060 ⇒ 00:39:37.780 Samuel Roberts: Alright, sounds.
387 00:39:37.780 ⇒ 00:39:38.110 Hannah Wang: Right.
388 00:39:38.110 ⇒ 00:39:39.330 Samuel Roberts: Have a good one. Bye.