I'm Pietro De Angeli

WIP

Reimplemented the paper Efficient Temporally-Aware DeepFake Detection using H.264 Motion Vectors. Matched the reported accuracy (71%) while reducing inference cost by replacing MTCNN with a lightweight YuNet-based face detector.

Extended the paper Layered Image Vectorization via Semantic Simplification by introducing radial gradients, achieving a +44% improvement in CLIP image score while reducing the average number of generated vector elements (-3% paths and -8% points).

Improved a GNN-based Max-k-Cut solver by hybridizing a Relax-Optimize-and-Sample (ROS) pipeline with Genetic Algorithms and Ant Colony Optimization. On 48 large-scale Gset instances (up to 20k nodes, k = 2, 3, 10), bio-inspired methods consistently outperform the learning-only baseline, with up to +1.24% average improvement and best performance on k = 10 (≈ 79% of instances).

Task-oriented travel planner built as a hybrid LLM dialogue system, handling flight, accommodation, and activity bookings through multi-turn conversations. Combines an LLM (Llama 3.1) for NLU/NLG with a schema-driven dialogue manager for controllable behavior — slot-filling with carryover across bookings, intent splitting for complex requests, and explicit confirmation before booking. Integrated with Amadeus APIs for live hotel and activity data.

Autonomous Agent for Deliveroo.js environment. The agent uses a BDI architecture with a PDDL planner to plan and execute actions in a dynamic environment. The agent is able to adapt to changes in the environment and achieve its goals. It can handle NL instructions using LLM interface and calling tools to solve 3 levels of tasks

Fine-tuned an English pretrained ASR model to Italian using NVIDIA NeMo, demonstrating effective cross-lingual transfer and achieving performance (CER: 4%) comparable to the official Italian QuartzNet model.

Designed a multilingual speaker diarization system using NVIDIA NeMo pretrained models, integrating ASR, and Speaker Identification for media analysis.

Developed a fake news detection classifier (logistic classification and Neural network) with OpenAI text embeddings. Reached an accuracy of 99% on a toy dataset.