This project explores multiple deep learning models: CNNs, ResNet50, and Vision Transformers, for classifying coins in images. We use advanced segmentation techniques, custom over-segmentation filtering, and data augmentation to improve generalization. Developed for the EPFL IAPR course, this system demonstrates competitive accuracy and robustness for automated coin recognition tasks.
This project builds a full-scale movie recommendation system using Apache Spark, incorporating data analytics, keyword-based filtering. Implemented on the MovieLens dataset, the system supports efficient data preprocessing, incremental rating updates, and personalized movie recommendations through LSH and predictive models.
This project explores the design and implementation of a decentralized DNS system using blockchain and a gossip-based peer protocol. The system supports secure domain registration, updates, transfers, and resolution with robust anti-entropy synchronization and Proof-of-Work consensus. The project evaluates network resilience, consensus reliability, and mining efficiency.
This project combines personality and emotion detection with machine learning to predict the number of ideas generated by entrepreneurial teams. We process multimodal data (transcripts and audio) to extract MBTI traits, emotional profiles, and speaker features, and use these to model team-level idea generation. This work advances behavioral modeling in early-stage startup evaluation.
A robotic system developed for real-time order collection in a restaurant environment, combining line-following, object detection, and GUI-based control. The robot uses image processing, multithreading, and Bluetooth communication to collect orders from tables and transmit them for analysis and optimization.
The graph similarity algorithm for matching image-text graph pairs was too slow, particularly in the pairwise comparison step
Group similar-length samples to minimize VRAM waste and stabilize throughput in NLP tasks.
PPO and GRPO training with models >7B caused OOM errors on A100 GPUs due to multiple full model replicas. This post details optimization strategies to fix it.
Improving distributed training speed using vLLM, Flash Attention, LoRA, gradient checkpointing, and stable checkpoint recovery across multi-node systems.
Efficiently mining structured text or graphs using GPT-4 APIs while staying under 2M TPM.
Correctly configuring AMP and autocast led to 2× faster training on NVIDIA GPUs.
Avoiding redundant tokenizer calls accelerated validation by up to 3× during fine-tuning.
Published in bioRxiv, 2025
We introduce a generative framework for constructing large-scale kinetic metabolic models through latent space exploration. By repurposing pretrained neural network generators across different physiological contexts, our method enables efficient and interpretable inference of kinetic parameters, facilitating targeted model design for diverse metabolic behaviors.
Published in EMNLP 2025 (Findings), 2025
We propose GRAD, a dynamic demonstration-based approach where an LLM model is trained to generate input-specific concise demonstrations. By tailoring demonstrations to each input, our method offers better contextual support than traditional RAG approaches.
This thesis introduces a multimodal framework for fine-grained text-to-image retrieval in remote sensing, combining global and structured representations through scene graphs and semantic region embeddings. It addresses the limitations of CLIP-style global embeddings by capturing spatial, semantic, and relational details critical for fine-grained reasoning.
Published:
We present a novel framework to interpret and control the latent spaces of generative neural network models for kinetic metabolic modeling. By perturbing structured latent spaces learned via REKINDLE or RENAISSANCE, our method generates new dynamic models with targeted properties such as specific response times, regulatory bottlenecks, or alternative physiologies, unlocking deeper insight and reusability across metabolic contexts.
Published:
We propose GRAD, a dynamic demonstration-based approach where an LLM model is trained to generate input-specific concise demonstrations. By tailoring demonstrations to each input, our method offers better contextual support than traditional RAG approaches.
We developed an educational chatbot built on the quantized Gemma 2 7B model, optimized with Direct Preference Optimization (DPO) and enhanced with Retrieval-Augmented Generation (RAG). By leveraging fine-tuning on student-generated preference data and incorporating relevant external documents, our system significantly improves accuracy in answering STEM multiple-choice questions, outperforming baseline models like Mistral and Llama2.
This project investigates how to accelerate motor skill acquisition in reinforcement learning using curriculum-based learning, dimensionality reduction, and policy distillation. Using the Myosuite Baoding balls task, we explore how expert policies can be transferred to novice agents via PCA-reduced feature and action spaces, offering an efficient alternative to prolonged training times.
This project was conducted at DISAL, EPFL. We explore trajectory generation for multi-robot navigation using neural networks. We propose a scalable alternative to Webots simulation by training models using graph neural network, reinforcement and imitation learning. The final approach produces accurate trajectories in a lane-based environment, balancing precision and efficiency in robotic control.
I led applied research and prototyping efforts in multimodal AI, focusing on cross-modal representation learning, graph-based embeddings, and neural search systems. I developed scalable pipelines to generate scene graphs from satellite imagery and knowledge graphs from textual data, and to align their graph embeddings in a shared representation space.
Developed a machine learning pipeline to forecast daily ad spend on Google Ads based on client-specific campaign data. Deployed a web backend for dynamic budget strategy adjustment, automated alerts, and integration with Azure Cloud infrastructure.
During my studies, I served as a teaching assistant for multiple courses, assisting in lectures, labs, and tutorials