We authored a paper on Second Order Jailbreak where we delved into the risks posed by malignant intelligent actors spreading their influence over networks of agents with varying intelligence and motivations. We run experiments on a multi-agent environment available here. You can see all the different conversations the agents had with each other on our playground website. We've presented our work at NeurIPS 2023. The current version of the paper can be found here. We are currently working on a follow-up paper, which will include a more detailed analysis of the risks and the strategies that can be used to mitigate them.

Co-created and leading a volunteer-driven organization focused on AI safety and alignment, organizing talks, round tables, seminars and bootcamps to foster knowledge exchange and raise awareness about AI risks. During which I:

• Successfully led the AI Safety Fundamentals from BlueDot Impact, implementing the AI Alignment Curriculum.
• Organized a 2-week AI Safety bootcamp in Sep. 2023, aiming to skill up 20 motivated participants and encourage them to work in the field of AI Safety. Delivered in-depth training on the technical aspects of Transformer architecture, including mechanistic interpretability (Induction and Indirect Object Identification circuits) and provided lessons on RL, RLHF and jailbreaking of LLMs.

In my role as a Data Officer at NCCR Catalysis (EPFL/ETHZ), I collaborate with numerous labs across Switzerland to standardize practices and optimize data management. I am developing a technique for uniform catalyst storage and sharing, and facilitating open science through AI model platforms. Additionally, I am co-authoring a manuscript detailing our research on ML-based detection of metallic atoms in microscope images and in the process of publish a second publication that will enable researchers to label microscope images seamlessly

Creating a melody streaming platform using NextJS, PostgreSQL (Supabase), and AWS services around a event-driven architecture. I developed a custom contrastive ML model used for 0-shot classification (mostly tagging and genre recognition) and semantic search. You can have a look at the webapp here.

Gave practice sessions to students (between 20 and 50 per session) for the following courses:

• Algorithms and data structures
• Discrete math and probability
• Numerical methods
• Signals and systems
• Python

Developed and deployed a Mask-RCNN model for satellite image analysis, encompassing detection, segmentation, and area computation. Integrated the model with location information obtained from OpenStreetMap for accurate image matching. The model is currently in production.

One year exchange at EPFL, where I undertook a master thesis titled "Point cloud compression for DNA based storage" and achieved a grade of 5.75/6. This research involved building an end-to-end point cloud compressor for quaternary based coding while adhering to DNA sequencing, storage, and synthesis constraints. In addition to my thesis, I completed the courses Cryptography and security, Information security and privacy, Decentralized systems and Seminar in advanced topics in machine learning. Took part in the Effective Altruism Lausanne and GNU Generation associations.

Master degree in data science with a specialization in cryptography and computer systems. Grade: Cum Laude.

Bachelor degree in engineering with a specialization in applied mathematics and computer science. Grade: Cum Laude.