R&D Scientist @ ItalAI | M.Sc. Computer Science @ Sapienza University

I am a researcher and engineer bridging the gap between theoretical deep learning and production-grade AI systems. My work focuses on Multimodal AI, Hypercomplex Models, and Embodied AI (Robotics), with a specific emphasis on making models highly efficient and adaptable.

Currently, I am building end-to-end spatial-audio interaction engines for embodied social robots at ItalAI and conducting research on Vision-Language-Action (VLA) models at PINLAB.

• Status: Polishing codebase
• Focus: Weight-Space Meta-Learning & Embodied AI.
• Contribution: Engineered a parameter-generation system that creates task-specific adapters (LoRA) from visual inputs. The framework enables zero-shot adaptation for large-scale robotic policies without test-time optimization.
• Code: WIZARD

• Status: Academic Project
• Focus: AI Safety & Machine Unlearning.
• Contribution: Engineered a framework to selectively erase restricted concepts (e.g., "kick", "fall") from 3D motion generation models. Adapted state-of-the-art unlearning methods (ESD, LoRA, UCE) for Skeleton-Aware Latent Diffusion (SALAD).
• Code: Motion-Unlearning-Evaluation

• Status: IJCNN 2025 (Oral Presentation)
• Focus: Hypercomplex Deep Learning & Generative Media.
• Contribution: Pioneered ResQu, a quaternion wavelet-conditioned diffusion model that dynamically adjusts conditioning strength across denoising stages. Surpassed existing SOTA (StableSR) by over 19% in PSNR on DRealSR.
• Code: ResQu

• Status: B.Sc. Thesis
• Focus: Reliability, Geometric Deep Learning, & Confidence Calibration.
• Contribution: Proposed a hyperbolic radius-based calibration method using Poincaré balls, successfully reducing Expected Calibration Error (ECE) by 50% on CIFAR-100.
• Code: Radius-Regularization

R&D Scientist @ ItalAI Labs (June 2025 – Present)

• Architected a hardware-embodied audio-spatial interaction engine, integrating real-time speaker diarization and directional tracking to enable seamless multiparty engagement for social robots.
• Engineered a persistent memory architecture utilizing dynamic per-persona context buffers to maintain conversational state and interaction history across long-horizon temporal sessions.

Generative AI Engineer @ F1 Consulting

• Architected a DAG-based multi-agent LLM orchestration system (8+ specialized agents) for Aeroporti di Roma to handle non-linear reasoning paths.
• Optimized LLM retrieval latency by 75% using semantic caching, embedding pruning, and efficient context-injection.
• Built real-time, interruptible voice pipelines integrating ASR and streamed TTS for seamless user interaction.

Software Engineer @ HCL Software (Worked full-time alongside B.Sc. degree)

• 🏆 Maverick Award Winner (Awarded to the top 4 out of 100+ engineers for exceptional impact).
• Engineered a scalable dense vector retrieval system and internal RAG architecture, reducing developer debugging time by 40%.
• Refactored legacy C/HLASM codebases, eliminating 30% of code duplication across critical systems.

Languages

AI & ML Frameworks

Cloud, HPC & DevOps

I am currently open to collaborations in Embodied AI and exploring elite PhD opportunities.

• LinkedIn: linkedin.com/in/christianbianchiit
• Email: ch.bianchi02@gmail.com