Project Learning Blog – Nadir Jouhri

From Dashboards to Signals: What Building Reddit Early Signals Taught Me
Where the idea came from
Our group agreed that everyone would pitch one idea and that we would then vote on the most interesting one. The Reddit idea was mine, and it came from pure curiosity. I had recently started following a few companies on the stock market, Reddit among them. While researching what the company actually does, I kept asking myself how Reddit could create value in the age of AI.
I also work as a working-student data engineer. Those two perspectives led to one question: why should users always have to go to a dashboard and do their own analysis? What if the insight came to them instead, as a proactive signal powered by AI and automation? That became our project: turning Reddit conversations into early warnings and opportunities for enterprises.
My role: building and testing the prototype
My main contribution was building and testing our prototype. It is an application with several screens: one where an organisation enters its context and strategic goals, a dashboard, an AI data consultant you can chat with, and an email generator that sends the key findings as a signal. The email is the core of the idea, because the user doesn’t have to go looking for anything.
I also made sure I stayed involved in every other part of the project. That mattered more than I expected. A prototype only works if it fits the story you are telling, so I had to understand our strategy, business model and positioning as well as the code.
The decision that shaped everything: which data to use
The hardest question was which data to use. I wanted to connect directly to the Reddit API, but our access request was rejected. I had to find another reliable source.
I turned to Hugging Face and connected to an existing dataset that met our requirements. That worked, but it created a new problem. Loading and storing the data was slow, and the dataset was so large that every question became token-heavy. Our platform is connected to my own Claude API, so my Claude had to work through that huge dataset every time, and that costs real money.
I then decided to write a Python script that generates our own dataset. The data is dynamic and changes every time the application starts, so it behaves more like a live feed. This gave me the balance I was looking for between data that is useful and plausible, token efficiency, and speed. The trade-off is honest: with synthetic data we can test the workflow and the quality of the AI’s reasoning, but not real-world precision.
What the evidence changed
Testing taught me more than building. In an early version, the AI consultant only received a small slice of the data as context. When we tested this on synthetic events, it rarely contained evidence of what had actually caused a trend. After I redesigned the context so the AI got an aggregated overview plus the relevant detail, the share of events where the AI could see the cause rose from about 49% to about 78%.
That surprised me. I had assumed more data would automatically give better answers. In practice, what matters is what you put in front of the model and how you structure it.
We also found things that did not work. The urgency scores did not discriminate well: they clustered in the upper range instead of clearly separating routine noise from real risk. In some answers, the AI consultant claimed a “direct correlation” that the data did not support. It also sometimes used facts about the company as if they were evidence. We did not hide these failures. We put them in our evaluation and used them to shape our recommendation.
Teamwork
The collaboration was generally positive. We sometimes struggled to schedule meetings because of different availabilities, but we found solutions so that everyone stayed involved. The best moments came when everyone was physically in the same room. We could discuss ideas, give each other direct feedback and even argue, and those sessions were where we made the most progress.
How my view of AI and strategy changed
Building the prototype made me realise how easy the technical part is to copy. Embeddings, clustering and LLM summaries are available to any competitor. What is hard to copy is Reddit’s long history, its community-specific baselines and the feedback from analysts who verify the signals. This is why our final recommendation was to build Early Signals but prove it on real data first, and to sell verified signals rather than raw data.
The project also showed me the strategic tensions. A product that sells insights could compete with the data licensees who already pay Reddit. And as customers start using their own AI agents, finding information becomes cheap, while trust in that information becomes the scarce good. That changed how I think about AI: the value is less in the model and more in the verification, the context and the trust around it.
Looking back
I loved working on this. It is exactly what I was looking for in my studies: thinking strategically about what technology makes possible, and then building it in the form of a prototype. I would have liked more time to develop it further, because I want to keep growing technically and stay close to what is possible today.
The biggest lesson is that technology has an enormous number of dimensions, which makes it complicated very quickly. The challenge for us as students, and soon as professionals, is to translate and simplify all of that so it creates real value.