Conf Companion

Framing the Problem

What is the problem you are trying to solve? Who does it affect?

---

Thursday someone asked me early in the morning which talks I am going to. I planned it out before my trip from Hamburg to Amsterdam but forgot which ones I wanted to go to. And after taking another look on the timetable, I wasn't so sure anymore what I picked in the first place. Also I did not have much time and had to make a quick decision, because the first talks were beginning in a few minutes! Yikes!

That's where the Conf Companion comes in.

Idea Explanation

What is your idea? How does it fix the problem?

---

My idea is a RAG AI chat interface, fed with the data of the conference websites. I can quickly ask the chat "Hey, can you recommend me some talks about Performance?" and the AI gives me some recommendations.

I can also ask "Who is Wes Bos?" or "Where is the venue located?"

Another feature is the timetable with the ability to set favorites.

To have all the data in one place, I also added data about the individual speaker.

I also designed a little mascot, inspired by good ol' Clippy: meet Lanny, the Lanyard. His friends also call him by his surname Badger.

Implementation

How do all the pieces fit together? Does your frontend make requests to your backend? Where does your database fit in?

---

So the data is scraped from both the conference websites – this data could be directly coming from the CMS in the future – and the data is then already chunked before the OpenAI embeddings model creates the vectors that fill up the PostgreSQL database hosted by Supabase with the pgvector plugin inclusive.

Before the frontend does requests through the Vercel AI SDK, the database gets searched for the top-k similar chunks and those chunks in combination with the question gets sent to OpenAI to generate the answer. There are no special boundaries or system prompts set.

The frontend is build upon TanStack Start and KendoReact.

I originally started to do a mockup with the Kendo Figma UI Kit but found myself to just hack away.

Challenges

What did you struggle with? How did you overcome it?

---

Setting up the database was a hassle at first – that's where AI came into play. The Supabase MCP was very helpful in getting all in place.
Also I am not the biggest fan of TailwindCSS. I thought I would be quicker with it but I always find myself to be...well....slower. I prefer CSS Modules and SCSS.

I also focused on implementing everything from Kendo UI myself, no AI implementation, no AI sparring, as I am currently working on my own frontend component library with a keen eye on accessibility. Going through how other companies structure their components and libraries lets me learn a lot.

Accomplishments

What did you learn? What did you accomplish?

---
I definitely learned much about Kendo UI and its capabilities. The amount of different components and their depth were surprising.

I also set up my first vectorized data pipeline and chat interface, hooray!

I also did accomplish to not sleep 😴 stay awake and focus on the task, despite having only this one night to work on the project.

Next Steps

What are the next steps for your project? How can you improve it?

---

For starters a connection to the actual data could be nice. One could improve on many aspects there: food plans with ingredients/allergens, what's available at the coffee spots, a digital floor plan/map, some interesting repos of the speakers – information the chat interface could be fed with.

A speech-to-text and text-to-speech option could also be quite nice. My first goal would be to make it as accessible as possible.

Some minor UI tweak could be made for now and a dark mode isn't implemented yet.

Down the line I could see a more homogeneous UI.

Notifications would be a useful feature – get notificated if something changed or one of your favorites is coming up!

I could also see conferences do Push-Messages through the application and let users notify about "Hey there! If you want some desserts, you can find them here and there!"

Actually my first idea (and half-finished project) was a digital floor planner: drawing the floor plan and its different levels on a canvas with rectangles, lines, circles, you name it. And then assign areas to them or if lines are walls etc. The ability to trace an image through AI or by hand. Or do the best of both worlds.

You were able to import this as GeoJSON data into the Kendo UI Map component and were able to look up routes on the plan from say "Sponsor Booth A" to "Meet-Up Area C". Sponsors could be highlighted on the map and before the conference the system could be set up to mark spots for potential sponsors.

I also prototyped a function for users to set a time restricted pin on the map for quick meet-ups with like-minded people:

breakout sessions after talks, find new friends, 'looking for groups' after the conf etc.

Others peoples pins could only be shown if the user wants them to.

I even tested some 'location on floor plan through wi-fi' options to pinpoint a users location in the building. Which would have paired well with the routing system to guide users to their location.

I think such a system has a lot of potential and would like to explore this further as such a system, when working reliably, could easily be transformed into a SaaS.

So: lots of ideas!

And also: Lanny mascot plushies.