Project: Large-scale Dataset and Foundation Models for Trajectory Prediction in Aviation

• Contributed to developing and open-sourcing a large-scale trajectory prediction benchmark for aviation GitHub and HuggingFace
• Developed parallelized Python processing pipeline for a large-scale trajectory dataset (~20 TB), reducing processing time by over 60% (from 40 to 15 hours) for scalable analysis and processing
• Assisted in training experiment sweeps of transformer models, achieving state-of-the-art performance on benchmark trajectory prediction datasets using Pytorch-Lightning
• Developed data-splitting algorithms to avoid data leakage and ensure deterministic splitting for fair reproducibility
• Co-authored and submitted a research paper to NeurIPS 2025 Datasets and Benchmarks Track, detailing the dataset, experiments, and model design

Project: Generative Robot Manipulation

• Engineered workflow automation tools for a robotic painting system that uses learned stroke-generation models to reproduce input images
• Designed and 3D-printed a calibrated workspace platform in Fusion 360, automating calibration and reducing setup time by over 70% (2 hours to 30 minutes)
• Built a full-stack GUI with Django, React, and Docker for seamless user-robot arm interaction
• Trained parameter-to-visual-output mapping models to improve stroke accuracy, motion expressiveness, and closed-loop feedback control

Project: Contact-rich Human-Robot Interactions using Propioception and Soft Actuation

• Designed and built a social robot prototype with an inflatable, soft outer layer enabling safe, contact-rich navigation and interaction with humans
• Integrated a ROS-based navigation stack using the CMU Autonomous Exploration tool, enabling autonomous navigation and mapping on a custom-built mobile robot
• Engineered and deployed a proprioception algorithm based on a 3D deformation classification algorithm using Open3D and PyTorch, allowing the robot to detect touch types and trigger context-specific responses
• Presented a demo of the prototype at the IEEE International Conference on Robotics and Automation (ICRA)

• Engineered and deployed a RAG-powered virtual assistant on AWS Bedrock, ECS, and ECR, and SerpAPI, enabling instant access to contextual portfolio data
• Designed and iteratively refined prompt-engineering workflows to optimize LLM outputs for risk scenario analysis, improving output accuracy and consistency by over 95% (5 hours to 10 minutes) and enhancing decision making
• Automated backend workflows and optimized data extraction from documents and web sources, with AWS EC2 and Azure AI, cutting forms completions by over 95% (15 minutes to 25 seconds)
• Developed scalable APIs with Flask, AWS ECR, and ECS for dynamic production workload handling
• Implemented deep learning models for document classification, image segmentation, and text extraction

• The objective is to create bots in order to automate Miniso's product placement
• The RPA proceses are created with Automation Anywhere
• The bots main requirements where to upkeep financial reports, item prices, quantities and internal data
• The automation sped up proceses from around 2 hours to 15 min and eliminated repetitive tasks done from end users

• The goal of the project was to detect alarm anomalies on telecommunication services provided by Orange France
• Used unsupervised algorithms for anomaly detection of alarm systems
• Evaluated unsupervised algorithms with statistical criteria in order to determine the best-one according to data set
• The models were implemented for testing phase
• The time that the supervision team took to detect certain anomalies was reduced