hello world!

Dexterous Manipulation via Embedded Posture Graphs of Hands with N-DoF

Soft Robotics Lab, ETH Zürich, 2026 - present. Supervisors: Chenyu Yang, Davide Liconti, Prof Dr Robert Katzschmann

Building a co-embedded graph of hand postures across end effectors with differing morphology and degrees of freedom, for generalized posture mapping. Our method captures posture connections and enables representation learning via a nonlinear manifold, learning a shared representation across hands.

The embedding acts as input to train a neural cross-embodiment retargeter between N-DoF hands, and as a compact posture representation for improved sample efficiency in RL-based and other downstream tasks.

NAICS-Aware Graph Neural Networks for Large-Scale POI Co-visitation Prediction

Intelmatix Labs, MIT, PSU, 2024 - 2025. Collaborators: Yazeed Alrubyli, Dr Omar Alomeir, Dr Abrar Wafa, Dr Hend Alrasheed, Dr Mohsen Bahrami

A novel GraphSAGE-based architecture that integrates learnable NAICS business taxonomy embeddings with multi-modal spatial-temporal features to predict population-level co-visitation patterns between 92,486 brands across 48 US states.

The model achieves a 157% improvement in R² over state-of-the-art baselines by learning business category relationships through end-to-end training on 45.3 million graph edges, enabling scalable edge regression on extremely sparse co-visitation networks.

Paper: arXiv:2507.19697

Energy-efficient Path Planning for Autonomous Drones in Inspection Tasks

Robotics & Perception Group, UZH, 2025. Supervisors: Leonard Bauersfeld, Prof Dr Davide Scaramuzza

Traditional sampling based path planning methods often struggle to accurately account for complex drone dynamics when evaluating the cost and feasibility of a path to a sampled point. Our work is based on 3D Dubins curve infused RRT*, which allows us to constrain drone dynamics through the help of 3D Dubins curved paths, while expanding the state space to include energy-related variables such as velocity and indirectly, the curvature of the turns, in order to minimize the energy expenditure of the drone for the planned path.

Stack: ROS Agilicious PythonRobotics

Knowledge Graph Retrieval-Based LLM Agent for Factual Querying

UZH, 2024. Collaborators: Raffael Botschen

A chat agent able to respond to natural language user queries based on factual information retrieved from free text documents represented in the form of multiple knowledge graphs. Relied on Llama API for prompt engineered query classification and for natural language response generation. The system generates SPARQL queries to accurately retrieve information from the knowledge graphs, and performs prior entity matching for node filtering. This approach enabled accurate, context-aware answers through structured knowledge graph retrieval while maintaining natural language interactions between the user and the agent on the frontend.

Stack: Ollama NumPy SciPy RDFLib SPARQL

Geospatial Site Selection Engine

Intelmatix, 2022 - present

A location intelligence engine that models the urban dynamics of a city to perform tasks such as real estate valuation or site selection for a new QSR branch. Relying on a graph-based saturation model, a branch cannibalization model, accessibility scoring for POIs and demographic groups, route optimization, and a catchment model, it can predict the future sales for unknown locations and recommend the highest-impact site for a new branch to open within the city.

The value prediction model receives inputs such as accessibility scores, market saturation, cannibalization score calculated based on the catchment of a location and its road network around it, and several demographic variables. This value prediction model can be trained to predict outputs such as future sales for a site, or the value of a land or real estate. Thanks to this flexibility on the target variable, the site selection engine can be scaled to several use cases that relate to urban and geospatial analytics.

Stack: PyTorch scikit-learn NumPy FastAPI Docker AWS S3

Unstructured Financial Data Extractor

Dealogic, 2021 - 2022

A data wrangling and extraction engine that relies on various NLP methods such as named-entity recognition, sentiment analysis, and text mining in order to extract highly accurate financial data from unstructured sources like free text, PDFs, and legal documents published by the US Stock Exchange Commission.

The goal of the project was to build a graph database to store the relationships of special purpose acquisition companies (SPACs), their acquired companies, their sponsors and sponsor affiliates, as well as all involved individuals' connections to these entities. In addition to capturing such relationships, we also wanted to collect additional information about the entities themselves, such as stock and warrant prices, ticker symbols and stock exchanges, purchase conditions and warrant expiry dates, etc.

The raw filings are collected and the data points are extracted near real-time, with minimal delay after publication on SEC. A NER model was trained on historical SEC documents, and it is able to process and extract data from various filing types such as 424B4, S10, quarterly and annual reports.

Temporal Risk Evaluator of Rising Household Debt

Stockholm University, 2019 - 2020

A risk assessment model that examines the effect of rising household debt on the financial markets in Sweden based on various time periods before, during and after the 2008 financial crisis.

For several years the household debt-to-GDP levels in Sweden have been close to that in the US prior to the 2008 financial crisis, coupled with a steady growth in housing prices. Considering the historical burst of the housing bubble in the US which triggered the crisis and had ripple effects on the financial markets, institutions such as the Swedish Central Bank and the National Institute of Economic Research in Sweden have been raising concerns.

This model was built to understand the market risk posed by the recent record high household debt-to-GDP levels, consisting mainly of mortgages, while controlling for variables such as consumer price index, repo rate, industrial production, oil prices, unemployment rate and household income. The financial markets were captured by major stock indices such as OMXS30, yield curves on long-term government bonds, and volatility indices like SIXVX.