Research on Hierarchical Graph Co-Clustering

March 2020 - August 2020

Cooperation with DIG group

Abstract: A well-known metric for quantifying the similarity between two clusterings is the adjusted mutual information. This paper proposes a novel adjustment based on pairwise label permutations, leading to a lower time complexity while maintaining quality. Experiments on synthetic and real data are presented for comparison.

The objective of this work was to develop a metric that assesses the quality of a tree in terms of graph reconstruction.

Mentors: Prof. Thomas Bonald

Paper Github

​ Text Classification with Deep Learning

December 2017 - June 2018

Abstract: Nowadays, retail e-commerce sales are quickly increasing. Large online e-commerce websites serve millions of users’ requests per day. Therefore it is necessary to make the processes of registrations and purchases as much convenient and fast as possible. Therefore, the problem of advertisement automatic category prediction is very important in terms of saving moderators’ time and as a result, decreasing the number of necessary moderators to process them. Effective algorithms which would work with text data, have high accuracy and appropriate speed are in high Demand.

The objective of this thesis was to build an effective model which would have high accuracy and appropriate speed for classification of advertisements at the ecommerce platform Jiji.ng. In particular:

• consider different models that are used for texts classification
• compare performance of Deep Learning arhitectures
• compare efficiency of CNNs vs RNNs for NLP related tasks

Mentors: Prof. Anton Maltsev

Thesis Presentation

AI networking glasses — winning hack at Project Europe

May 2026

Winner — Project Europe hackathon

Built in 24 hours and awarded first place at the Project Europe hackathon: AI glasses that let you walk into any event already knowing who to talk to. As you move around naturally, the glasses recognise faces in the room, surface the few details that make a conversation relevant, and whisper a smart opener in your ear — turning random networking into meaningful conversations.

The end-to-end experience:

• Capture from Meta glasses — live video is streamed from Meta Ray-Ban glasses via the Meta developer SDK.
• Recognise faces in the room — a face-recognition model matches people to the profiles you want to meet.
• Enrich in real time — public signals from LinkedIn, X and blogs surface the few details that make a conversation relevant.
• Guide the opening — a short audio prompt hands you a natural opener so you stay present while the research happens in the background.
• Capture the conversation — speech-to-text transcribes and summarises each interaction into a searchable, follow-up-ready record.

LinkedIn Website

Leleka Lens — human data for robot learning

January 2026

A "Coursera for Robots": a global platform for collecting first-person human data for robot learning. People record everyday tasks — cooking, repairing, assembling, sorting — using smart glasses or a head-mounted phone, anonymise the footage on-device, and turn it into structured, privacy-preserving datasets that can train the next generation of robots, world models, and embodied AI.

Real-world robot data is scarce, narrow, and expensive to collect, so robots struggle to generalise across homes, objects, and environments. Leleka Lens turns everyday human experience into a living dataset for robotics, and gives contributors a creator economy — they build thematic channels of hands-on footage and earn when their data is licensed or used in research.

Website

Human pulse measurement with web camera

April 2015

University project

Our pulse is the rate at which heart beats. As our heart pumps blood through the body, we can feel pulsing in some of the blood vessels close to the skin`s surface, such as in our head, neck or upper arm. To the human eye, no matter how long and hard you stare at your wrist or someone else’s face, you would struggle to detect a change in color. For a computer, however, the tiniest per-pixel fluctuations are easy to detect. As our heart pumps blood around body arteries swell with bright red blood, which changes the color of skin slightly. Thus we can measure accurately our heart rate. This project works by combining several techniques common in the field of computer vision.