📍 Torino, Italy · 🌍 Open to Relocation (EU)
🎓 M.Sc. in AI & Data Analytics (4.0 GPA, 110L) @ Politecnico di Torino
📫 posta.fedcar@gmail.com · LinkedIn · GitHub

• Histomorph — WSI Virtual Staining with DiTs PyTorch Accelerate HPC Multi-GPU Work in Progress
  • Scaled distributed training of Diffusion Transformers (DiTs) conditioned on histopathology ViT foundation models across a multi-GPU HPC cluster.
  • Commissioned by academic faculty to extend previous coursework into an end-to-end WSI-to-WSI translation framework targeting formal publication.
  • Scaled data pipelines to 100k+ pathology tiles, optimizing custom high-throughput dataloaders and cross-node gradient synchronization.
  • (Manuscript in preparation for venue submission · Private draft & architecture available upon request)

• Speculative Decoding on MoE Mistral-7B Mixtral-8x7B MoE QLoRA
  • Implemented speculative sampling pairing Mistral-7B (Draft) with Mixtral-8x7B (Target) using a custom CE + KL divergence objective.
  • Achieved 1.45x wall-clock speedup on GSM8K and HumanEval while profiling memory bandwidth bottlenecks and MoE routing overhead.
• Edge LLM Inference Engine from Scratch Rust Candle Inference
  • Built an end-to-end LLaMA inference engine in pure Rust from scratch (RoPE attention, RMSNorm, SwiGLU) running on constrained edge devices.ù

• Resolving Observation Lag in Multi-Step Conformal Prediction PyTorch NexCP Theory & Systems Manuscript in Review

  • Formulated an online, asynchronous calibration framework using a Matrix-Based Pending Queue with In-Place Buffer Propagation to resolve the multi-step ($H$-step ahead) observation lag without lookahead bias.
  • Derived explicit finite-sample per-step marginal coverage guarantees grounded in the Non-Exchangeable Conformal Prediction (NexCP) theorem under temporal non-stationarity.
  • Demonstrated consistent marginal coverage stability across classical CQR, time-decay NexCP, and state-dependent ResCQR over standard time-series benchmarks (ETTh1/2, Traffic, Weather).
  • (Manuscript under review · Full paper draft available upon request)

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

  • Comprehensive extension of Reservoir Conformal Prediction (ResCP) integrating topological ESN memories, dynamic ESS-based adaptive blending, and structural safety floors against interval collapse.
  • Developed a streaming post-hoc risk controller (inspired by CRC and ACI feedback loops) that dynamically adjusts quantiles to satisfy asymmetric business cost targets during extreme concept drift (COVID-19 shock).
  • Scaled benchmark across 26 real-world energy datasets and deep backbones (iTransformer, TSMixer, TiDE), achieving a +15.4% mean IWS improvement over uncalibrated baselines and outperforming standard CQR in 51/78 configurations.
  • (Full thesis text & experimental codebase available upon request)

• DORAEMON calibration PyTorch Deep RL MuJoCo Sim-to-Real Paper Included
  • Formulated a two-phase Sim-to-Real framework decoupling robust generalized pre-training from gradient-free test-time adaptation for robotic locomotion (Hopper-v4, Walker2d-v4).
  • Trained contextual policies ($\pi_\theta(a|s, \xi)$) under a constrained entropy-maximization curriculum (DORAEMON) with analytical Beta distributions and KL trust-region constraints to mitigate catastrophic forgetting.
  • Implemented test-time System Identification via the Cross-Entropy Method (CEM) to recover hidden physical parameters ($\xi$) without online gradient updates, resolving dynamic aliasing (mass-gravity disentanglement).
  • Calibrated policies recovered near-optimal performance in extreme outlier environments where static baselines collapsed (e.g., reward improvement from ~1200 to ~2600 in Walker2d-v4).

• GridShift — Huawei Tech Arena PyTorch Tree Models Optuna Competition
  • Engineering an end-to-end multi-horizon risk forecasting pipeline ($P \in [0, 1]$) for AI Data Centers (AIDC) under extreme weather events and 3 incoming-feeder topologies.
  • Formulated custom Compound Poisson-Gamma Tweedie NLL objectives ($p \in (1,2)$) and fault-window chunking to solve severe zero-inflation on regional outage spikes.

• Remote Filesystem Rust FUSE TypeScript Systems Programming
  • High-performance remote filesystem implementation leveraging FUSE in pure Rust. Handles low-level POSIX filesystem operations, asynchronous I/O caching, and robust network RPC synchronization.

• 🥇 Global Winner — GitHub "For the Love of Code" Hackathon (2025): Built Jukebox CLI, a high-performance terminal audio engine written in Rust.
• 🎓 Academic Excellence: Double 110/110 Cum Laude (B.Sc. in Computer Engineering & M.Sc. in AI), 4.0/4.0 GPA (29.9/30 avg).

_{Note: Manuscripts under review, competition repositories, and private implementations are available upon request for interview and evaluation purposes.}