Umberto Puddu

Computer Engineering & Mathematics @ Purdue University

Portfolio · LinkedIn · ResearchGate · upuddu@outlook.com

Interning this summer — open to opportunities from Fall 2026 onward.

I build intelligent systems and the interfaces that make them feel alive — transformer models and numerical libraries written from the math up, embedded firmware, and real-time simulations. My work runs from machine-learning research to systems programming to the occasional soldering iron. Right now I'm building AI applications at Reply.

Shipped

• EqoCalc — a published iOS app (Swift) for calculating pipe-fluid pressure drops. Live on the App Store.

AI & machine learning

• Continuous Attention Transformer — a research prototype recasting attention as continuous-depth neural-ODE blocks with spectral, Sobolev-smoothed kernels.
• Neural machine translation — a Transformer (Italian → English) implemented from scratch in PyTorch.
• neural-net-c — a multilayer perceptron with backpropagation and SGD in pure C99, no dependencies.

Systems & tooling

• Purdue Connect — a CLI that makes Purdue campus services — courses, dining, events — scriptable from the terminal.
• boiler-schedule — a course-schedule builder with conflict detection and PDF/SVG export; Go API backend, React frontend.
• HKN-POS — an IMAP → PDF-parsing → FastAPI webhook pipeline that turns order-confirmation emails into structured data.

Embedded & simulation

• PioNeer — handheld gaming console firmware for the RP2350; an LVGL launcher that side-loads games from an SD card.
• SlitDiffraction — a real-time Huygens–Fresnel wave-field simulator for macOS (Objective-C++ / Cocoa).

Languages

ML & data

Agentic AI

Web, backend & systems

_{Chicago, IL · upuddu@outlook.com}