I build reliable, scalable machine-learning systems—from distributed generative models and LLM inference to uncertainty-aware forecasting under distribution shift.

📍 Torino, Italy · 🌍 Open to relocation across the EU
🎓 M.Sc. AI & Data Analytics, Politecnico di Torino — 110/110 cum laude · 4.0/4.0 GPA
📫 Email · LinkedIn · GitHub

Speculative Decoding on MoE · Mistral-7B Mixtral-8×7B MoE QLoRA

• Implemented speculative decoding with Mistral-7B as draft model and Mixtral-8×7B as target, using a custom cross-entropy + KL-divergence objective.
• Achieved 1.45× end-to-end wall-clock speedup at (k = 2) on GSM8K and HumanEval; profiled memory-bandwidth bottlenecks and MoE routing overhead.
• See the repository for benchmark configuration, hardware context, and limitations.

Custom LLM Inference Engine · Rust Candle CUDA

• Implemented the LLaMA forward pass in Rust—including RoPE, RMSNorm, attention, and concurrent REST serving—without high-level LLM-inference wrappers.
• Designed a thread-safe, preallocated KV cache with in-place updates to remove decode-time allocation overhead.
• Achieved up to 14× decoding improvement under documented CUDA benchmark settings.

Resolving Observation Lag in Multi-Step Conformal Prediction · PyTorch NexCP Theory & Systems

• Designed an online asynchronous calibration framework that handles delayed (H)-step-ahead observations without look-ahead bias.
• Derived finite-sample, horizon-specific marginal-coverage guarantees grounded in Non-Exchangeable Conformal Prediction under temporal non-stationarity.
• Evaluated coverage stability against CQR, time-decay NexCP, and state-dependent ResCQR on ETTh1/2, Traffic, and Weather benchmarks.
• Completed research manuscript; preparing public preprint and venue submission. Technical materials available upon request where permitted.

Adaptive Conformal Calibration for Non-Stationary Multi-Step Forecasting · M.Sc. Thesis Darts PyTorch

• Extended Reservoir Conformal Prediction with topological ESN memory, effective-sample-size-based adaptive blending, and safeguards against interval collapse.
• Built a streaming post-hoc risk controller, inspired by CRC and ACI feedback loops, for asymmetric business costs during extreme concept drift.
• Evaluated 26 real-world energy datasets with TiDE, TSMixer, and iTransformer backbones; achieved +15.4% mean IWS over uncalibrated baselines and outperformed standard CQR in 51/78 configurations.
• Completed thesis research; manuscript preparation in progress.

Histomorph — Whole-Slide-Image Virtual Staining with DiTs · PyTorch Accelerate ViT HPC

• Developed an end-to-end WSI-to-WSI virtual-staining research pipeline using Diffusion Transformers conditioned on histopathology ViT foundation models.
• Scaled training across a multi-GPU HPC cluster and built high-throughput data pipelines for 100k+ pathology tiles, including custom loaders and distributed synchronization.
• Built an out-of-core pathology preprocessing workflow and released a PyPI package for WSI tile processing.
• Core system complete; manuscript and public release preparation in progress. Architecture and permitted materials available upon request.

Nixtla / NeuralForecast Contributor · PyTorch Time-Series Foundation Models

• Extended the upstream iTransformer implementation with native support for future exogenous covariates in PR #1607.
• Delivered production-grade PyTorch code and unit tests through the project’s upstream review process.

GRPO Alignment Pipeline for LLM Reasoning · PyTorch vLLM DeepSpeed Accelerate SLURM

• Developing a config-driven distributed GRPO pipeline for Qwen2.5 models on mathematical-reasoning tasks.
• Integrating high-throughput co-located rollout generation through vLLM with multi-GPU training via Accelerate, DeepSpeed ZeRO-2, and DDP.
• Building automated reasoning evaluation with unbiased pass@k estimators on MATH-500, OlympiadBench, and Minerva.
• Research implementation is ongoing; technical materials available upon request.