Serverless ML Projects at KTH - Real-World Applications

We're diving into the new year with our first meetup!

Keeping up with last year's 'tradition,' we've got talks lined up from various student groups at KTH. The students will present and demo their projects - they have built complete ML systems, solving important prediction problems on novel data sources, using only serverless technologies.

Agenda:

17:30 - 18:00: Doors open
18:00 - 18:05: Welcome
18:05 - 18:20: Scraping 1.3 million home prices - and getting away with it
18:20 - 18:35: BYOSC Build Your Own Scalable Chatbots
18:35 - 19:05: Pizza & Beers
19:05 - 19:20: Transitmap
19:20 - 19:35: DeLight - The Delay Flight Predictor
19:35 - 19:50: Where was this image taken? - Visual Geolocalization in the era of Deep Learning
19:50 - 20:45: Networking

–

Presentations:

Scraping 1.3 million home prices - and getting away with it
Pierre Le Fevre - KTH Student
Emil Karlsson - KTH Student

Hempriser is a property price prediction system for Swedish properties listed on Hemnet, developed by Emil Karlsson and Pierre Le Fevre at KTH. The system, born from the absence of a public API from Hemnet, involves intricate web scraping, data cleaning, and machine learning model training to predict property prices.

BYOSC Build Your Own Scalable Chatbots
Matteo Circa - KTH Student
Federico Giarrè - KTH Student

When preparing for university exams, having a partner has been proven to be essential to discover knowledge gaps and clarifying specific doubts on the topic treated during classes. While chatbots based on LLMs such as ChatGPT, Phind and Clod are providing help to students already, they cannot provide a lecture/material-specific help on the students’ university courses. We propose to create a system to fine-tune chatbots on specific material of specific courses. Thanks to this, we will create study buddies for the courses of a typical university student, able to answer doubts, generate questions and more.

Transitmap
Fabian Zeiher - KTH Student
Jonathan Arns - KTH Student

Transitmap is a real-time visualisation of all of Sweden’s public transport. Transitmap displays a map with the current locations of buses, trains, and ferries. The user can follow the movement of vehicles with less than 5 seconds delay between real-world and digital representation. Transitmap was born as a university project for a big-data processing course. This presentation will focus on a recent extension of the project with a neural network based delay prediction feature for the metros in Stockholm.

DeLight - The Delay Flight Predictor
Giovanni Manfredi - KTH Student
Sebastiano Meneghin - KTH Student

DeLight consists of a Serverless Machine Learning Pipeline, that allows a ML model trained to daily predict the departure delay of all the flights departing from Stockholm Arlanda Airport, for the following two days. A new model is trained daily, adding to the initial historical data extracted from various API, covering flights and associated weather conditions from the full 2023 year, and new samples extracted on a daily-basis. The model performances, as well as the delay predictions, are accessible through a friendly web interface on Hugging Face.

Where was this image taken? - Visual Geolocalization in the era of Deep Learning.
Adrian Häusler Redhe - KTH Student

Have you ever played Geoguessr? The game where you get dropped somewhere in the world, where Google Street View is available. Your goal? To place yourself on a map using visual cues only.

Discover the process of creating a simple ML-system designed to do Visual Geolocalization in a city-level setting, complete from data collection to evaluation.

–

About the event

Date: February 22nd, 17:30 - 20:30
Location: Hopsworks Office (Åsögatan 119, Plan 2, 116 24 Stockholm)
The venue this time is at the Hopsworks Office. As the office is sometimes difficult to locate we have made this map for everyone to follow.
Directions: 2-minute walk from Medborgarplatsen.
Tickets: Sign up required. Anyone who is not on the list will not get in. The event is free of charge.
Capacity: Space is limited to 60 participants. If you are signed up but unable to attend, please change your RSVP by February 21st.
Food and drinks: Food and drinks will be provided.
Questions: Please contact the meetup organizers.

–

Code of Conduct
The NumFOCUS Code of Conduct applies to this event; please familiarize yourself with it before attending. If you have any questions or concerns regarding the Code of Conduct, please contact the organizers.