Building an AI-Powered Social Sentiment Agent | Nerd Stuff and AI with Humanitarians AI
This walkthrough builds the Mycroft Social Sentiment Agent end to end in n8n, using Groq's fast Llama 3.1 inference to turn developer chatter into scored investment intelligence.
Developer forums produce a constant stream of small signals about where technology is heading, a Stack Overflow question about a new framework, a GitHub repository jumping from a hundred stars to a thousand overnight, a Reddit thread about a tool nobody's heard of yet. Individually these are noise. Read enough of them and patterns emerge, but no one has time to sit and read the entire internet. This build tackles that problem directly: an automated pipeline called the Mycroft Social Sentiment Agent that reads developer discussions across three platforms and turns them into structured investment intelligence.
The problem it's solving
The starting point is simple to state and hard to do manually. Stack Overflow is full of developers stuck on problems or announcing breakthroughs, each one a tiny signal about adoption. GitHub's trending page shows star counts jumping overnight as developer attention shifts toward a project, another market signal. Both are genuinely informative and both are effectively untrackable by a human scanning pages one at a time. The system built here automates that reading and scoring process end to end.
Three collectors, one trigger
The workflow starts with a manual trigger, functioning as a single "go" button for the whole intelligence-gathering run. Firing it launches three data collectors simultaneously. The Stack Overflow collector calls Stack Overflow's API asking specifically for the latest questions tagged artificial intelligence, filtered to fresh content from the last few hours rather than older archived questions. Parallel nodes handle GitHub and Reddit the same way. Each collector node extracts what actually matters from the raw response, timestamps, authors, and content, before all three streams meet in a single "combine all data" node. At that point, everything is fresh content pulled from across the internet, but it's still raw, unscored data.
Groq, not Grok: the sentiment engine
The intelligence layer runs on Groq, spelled G-R-O-Q, which the build is careful to distinguish from Grok, spelled with a K, the unrelated AI model from Elon Musk's company. Groq is an inference infrastructure company that runs open-source models, including Meta's Llama, at very high speed, reported in the video as roughly ten times faster than more traditional providers. Getting access is a matter of creating an account at Groq's console and generating an API key from the API Keys section, which comes with what the video describes as a generous free tier, enough for thousands of calls without cost in the presenter's own experience.
Inside the workflow, the sentiment node uses a Llama 3.1 8-billion-parameter model with a deliberately narrow prompt: analyze the sentiment of this text and respond with only positive, negative, or neutral, no explanations, just a clean signal. A follow-up node processes that response and attaches a confidence score, because the model isn't always certain, and for a sentiment pipeline it matters to know when the underlying data is shaky rather than treating every output as equally reliable.
A second pass for topic, then a filter for quality
A second call to Groq, using the same Llama 3.1 8B model, handles topic classification: is the discussion about LLM development, AI hardware, or an investment-relevant conversation. Running sentiment and topic as two separate passes gives the system real context. A phrase like "Nvidia memory workflow" might read as negative in isolation, but it's actually describing a technical issue, not market sentiment; "Nvidia losing market share" is a genuine investment signal. Separating what's being said from how positive or negative it sounds is what keeps the pipeline from confusing technical complaints with market-moving news.
From there, a topic clustering engine scans for specific technical terms, API, SDK, neural network, GPU, and similar language, indicating developers are actually building something with future revenue potential, while also tracking engagement signals like GitHub mentions and Stack Overflow references. The signal-to-noise filter that follows is where most of the raw volume gets discarded: since social content is described as roughly 90% noise, every piece of content is scored from 0 to 20. Stack Overflow content gets a base 4 points, GitHub gets 3, and a random Reddit comment gets 2, with additional points layered on for length, technical depth, and engagement. Only items scoring around 8 or higher pass through, which is the mechanism that turns a large volume of raw content into a much smaller set of high-quality signals.
From signals to a stored dashboard
The surviving high-quality signals feed a data aggregation node that goes beyond simple positive/negative counting, producing outputs like an overall bearish or bullish read, a technical development activity level indicating whether developers are actually building, and trending topics showing what the community is focused on and why. A data export node then assembles this into what's described as an executive dashboard, covering sentiment broken down by platform, quality metrics, and trend analysis.
Storage runs through Google Cloud: a Google Cloud project is set up with an OAuth2 client, whose client ID and client secret authenticate the connection between n8n and Google Sheets. Once that connection is established and the account is authorized inside n8n, every workflow run writes data securely into a spreadsheet, with 25 different fields captured per high-quality signal, including timestamp, platform, author, and content, building an ongoing intelligence database that lets patterns emerge over time rather than in any single run.
Key takeaways
- The Mycroft Social Sentiment Agent pulls fresh, tagged content from Stack Overflow, GitHub, and Reddit simultaneously from a single manual trigger.
- Sentiment scoring runs on Groq (not to be confused with Grok), using a Llama 3.1 8B model with a deliberately narrow positive/negative/neutral prompt.
- A separate topic-classification pass on the same model distinguishes technical complaints from genuine investment signals, like "Nvidia losing market share" versus a routine "memory workflow" bug report.
- A signal-to-noise filter scores every item 0 to 20, using platform base points plus length, technical depth, and engagement, passing through only items scoring around 8 or higher.
- Results are aggregated into a dashboard-style output (sentiment breakdown, activity level, trending topics) and written to Google Sheets via an OAuth2-authenticated connection, capturing 25 fields per signal.
- The entire system uses free-tier APIs and is open source.
Try it yourself
The full build, including code, prompts, and setup details, is available in the Humanitarians AI GitHub repository under the Mycroft project. This is a hands-on project from Humanitarians AI's Nerd Stuff and AI series, aimed at making investment intelligence tooling transparent and buildable by anyone rather than locked inside a proprietary platform.
Full transcript(auto-generated, with timestamps)
[0:06]Hi everyone, I am Vin Rajapu and today I am going to show you something I built that reads the entire internet pulse on artificial intelligent investments. This is the Microsoft source sentiment agent an A powered system that analyzes thousands of developer discussions, trending projects and technical conversations to generate real investment in. So let me show you the problem we are solving. Here is the stack overflow right now. So look at all these questions. Some developer is stuck in some problem and the just discover a breakthrough. So each of these is a tiny signal about a adoption. But who has time to read all of these now check the GitHub's trending
[0:49]Page. So see all these star counts when a project jumps from 100 to thousand stars overnight. Thus developers outing with their attention. So that's a market signal but again impossible to track manually. So this is why I built an automated intelligence system. Let me walk you through exactly how it works. So everything start with manual trigger. Think of it as go button for our intelligence gathering. When I click this three different data collectors spring into action simultaneously. So first up stack overflow. So this little node is actually pretty smart. It is calling stack overflows API asking for the latest questions tagged artificial intelligence. Not old stuff just fresh questions from
[1:41]The last few hours. So now watch this process stackleflow node. So it takes the raw data and extracts what we actually care about. So the timestamps, the authors, the contents and everything that are necessary for valuable information. So same thing happens with GitHub and other node known as the which colors information which is Reddit. So all these three streams meet here in the combine all data node. So this is all a fresh content from across the internet but this is just still raw data. Now comes the intelligence part. So okay this is where things get really cool. See this note it says Grock LLM sentiment. Now quick clarification
[2:39]Because people always ask about this. So Grock spelled like G R O Q. It is not the same as Ellen Mos Gro with a K. Totally different companies. Gro is the credible a infrastructure company that runs opensource models like Metas Lama at absolutely insane speeds. We are talking 10 times faster than traditional providers. So let us go to this actual website Grock. So let me show you how uh it is set up. It is a gro console. Swaper clean interface, right? So, click on API keys here. See, uh I have already got one created. That's what connects our workflow to Groa. Want to create your own? Just
[3:30]Create the hit create API key. Give it a name like uh some asent and boom, you get your unique key. Copy this and keep it safe and you have got access to lightning fast AI. Best part, Grock has a really generous free tire. I have made thousand of calls and have not paid a penny. So now let's go back to our show sentiment workflow. So in this workflow, so let us click on this node. So here is what API keys goes. So see the model we are using llama 8b 8192 that's metas llama's 3.1 with 8 billion parameters and look at our prompt super simple analyze the
[4:22]Sentiment of this text and respond with only positive negative and neutral no yeses no explanations just clean signals. So let us go to next node. So this is process gro sentiment node. This node is crucial because it takes Grock's response and adds our own confidence scoring because here is the thing sometimes the AI is not sure and for sentiment distance we need to know when our data is shaky. Let us go to the next node. Here comes the Gro topic classification. So now here we get clearer and clever. We hit GROs APIs again and the same model. See we have using the llama 8 8
[5:17]Billion parameter model only. So this time we ask what are they talking about? Is it LLM development AI hardware investment discussion? So this double pass gives us incredible context. Nvidia memory warflow might sound negative but it's just a technical issue. Nvidia lossing market share. Now that's an investment signal. So this is how topic classification works. Now comes the part where we separate gold from garbage. I'm clicking on this topic clustering engine. So this is my favorite part. The topic clustering engine. This is where we get surgical about what actually matters. So look at what we are doing here. We are scanning for technical words like
[6:08]EAPI. So he here we are searching for uh API, SDK, neural uh neural network and GPU and coded and all those technical terms they're actually building something that's a future revenue. So we are also tracking developer engagement how many GitHub mentions stack overflow references when developers are exited products succeeded or not. So when they are frustrated whether company struggled or not these are all things are covered in this clustering engine component. Next we are moving to signal to noise filter. Now this is where magic happens. The social media is 90% garbage. Right? This node fix that every piece of content gets scored from 0 to 20.
[7:17]Tag overflow it will get automatic four points because those are real developers and GitHub gets three points. These are base points and uh random Reddit comment will get two points. Then we add points for length, technical depth, engagement and the things which only score about eight pass through. So this is where the filtering smart filtering takes place. So look at this. We started with multiple items. Now uh we are just down to high quality signals. See uh we only got high quality signals in the right. So that's the difference between information and intelligence. Next let us go to data traditions. So this is where everything comes
[8:12]Together. Sources sentiment output. So we are not just counting positive and negative anymore. Look at how we got outputs like here bearish bullish we will get uh technical develop acuity level are developers actually building. So trending topics what's everyone obsession with and why. So all these will we get covered in this particular node. So see here we get a bearish and technical community activity got high for this one and signal quantity assessment got strong for this uh for this input. Moving forward and data export node creates our executive dashboard. It's everything an investor needs in one view. So sentiment breaks down by platform quality metrics
[9:12]And trend analysis. So the last and final part is setting up for Google cloud integration. Now here is where things get a bit technical but I will walk you through the essential setup to store our intelligence data in Google Sheets. We need to authenticate with the Google's API using Google cloud platform. So let us go to Google cloud platform. I have already set up a Google cloud project called Microoft in Nate and integration. Here you can see my O 2.4 client. This is what connects N to Google sheet securely. See this client I created client ID that uh Google generates for authentication. This is the bridging between our
[10:06]Workflow and Google services. So let us go to workflow. So back in any item here is where the magic happens. I have already connected my Google account. See account created. Let us er this. So see here my account is already connected. So this green check mark says that uh I have used the client ID and client secret from my Google cloud project. So look at the client ID and the client secret below is what make authentication secure. Let us go back. So, so the beautiful thing is once uh the thing is set up, it just works. Every times our workflow runs, the data flows securely into Google Sheets
[11:16]Through this authenticated connection. So, here is something most people miss. Intelligence without memory is worthless. So every high quality signal we identified gets 25 different matrix attached. So what to show the so here the data flow directly into sheet one automatically. So it is simple but powerful. Check this out. Here's our actual intelligence database. timestamp, platform, author, content and like this we have 25 types of columns. So we can literally see market mode shifting before prices move. All right. Uh let's run these things live. Watch what happens. So let us click this execute workflow button.
[12:28]So boom see the green checks appearing. So that's success at each step. The numbers show items flowing through stack overflow items pulled GitHub items pulled and red ML discussions process. So all ming now gro is analyzing sentiment and classifying topics filtering noise and finally push into Google she's through our GCP authentication. So done it's just a couple of minutes we have analyzed real discussions understood the sentiment classified the topics and filtered the noise and stored everything in our intelligent database. That's Microsoft source sentiment agent. So it's processing real data from the technical communities that actually build a technology stack overflow developers, GitHub projects and machine
[13:16]Learning researchers. So entire thing is open source and it uses free APIs and generates real investment intelligence. So this is how we democratize in investment intelligence with the transparent understandable systems anyone can build and improve. So thank you for watching and happy building. Goodbye.
More videos
2:08Bridging the Pixel Gap in Browser Automation.
2:23How One Narrow Safety Rule Can Make an AI Less Safe Everywhere Else.
2:04Why splitting a chunk from its document makes it retrieve for the wrong question
4:20Three You Can Take Back. One You Can't.
2:21Why a 50-turn agent pays for the same screenshot 35 times unless it caches the pixels
1:53