Projects

Face-only age & gender estimation reaching MAE 3.58 yr on LAGENDA 84k from a single face crop — beating body-aware MiVOLO v2 (3.86) using public training data only. Built on a DINOv3-ViT-L/16 backbone with CORAL ordinal regression and four-phase progressive fine-tuning. Ranks #1 among non-LLM methods on the eight-dataset FaceAge benchmark.

Full-stack face recognition with >99.4% accuracy, 30 FPS, and >10,000 identifiers. Co-developed with Oxford University advisers. Deployed in hospitality and regulated gaming sectors.

ML re-ranking layer trained on billions of user signals, resulting in a ~6% revenue increase (~$30K/month). Full MLOps automation from data ingestion to rollout.

Developing efficient memory management mechanisms to eliminate bottlenecks and enable LLMs / LVMs to run reliably on NVIDIA Jetson and similar edge hardware, in collaboration with Dr. Ismet Dagli.

Voice AI product with real-time transcription, speaker diarization, content summarization, and semantic search — built on LLMs and state-of-the-art audio processing research.

Adapted Stable Diffusion to generate commercial-grade stock photography via fine-tuning and LoRA. Unified app supporting txt2img, img2img, inpaint, outpaint, upscale, and variation.

A growing suite of AI agents that automate repetitive daily engineering and research tasks — turning hours of manual work into seconds of autonomous execution. Agents handle automated code review & PR summarisation, literature monitoring (daily ArXiv digest), meeting preparation, data pipeline health checks, and report drafting from raw experiment logs. Built with multi-model orchestration using Claude Code and Gemini, observability via Langfuse, and LLM security guardrails.

Multi-label classification model for 25,000+ visual categories trained on 5M+ samples. Evolved from ResNest → Meshed-Memory Transformer → pure Vision Transformer.

Multi-task model predicting age and emotion using context-aware recognition to fuse facial and scene-level features on a noisy, in-the-wild dataset.

Clustered hundreds of thousands of unlabeled face images for automatic identity labeling, overcoming extreme class count, attribute variety, and label noise.

The first voice bot in the energy sector in Italy, built for Sorgenia at Pi School, Rome using Dialogflow and custom NLU pipelines.

Core NLP platform covering speaker identification, speech recognition, wake-word detection, and multi-domain dialog. Piloted at a healthcare center for elderly residents in Kyoto, Japan.