# New Potential Automatic Parallelizer

> Source: <https://deshi.ch>
> Published: 2026-09-16 22:47:17+00:00

[Filippo De Simoni](https://github.com/mrdeshi)

a glimpse into my Mathematics & Computer Science

[Scroll](#about)

About

## MSE student, Computer Science

Passionate about mathematics and algorithms in all their forms: from linear algebra in computer graphics to discrete math, optimization, and operations research.

Domains

- Algorithms
- Optimization
- Computer Graphics
- Machine Learning
- Discrete Math
- Systems Programming

Experience

1. 2023 - now#### Co-Founder & Backend DeveloperOutzone, Lugano Server-side architecture, REST APIs, and data management
2. 2024 - now#### Peer Tutor (Formal Collaborator)SUPSI, Lugano Tutoring in CS and Mathematics under collaborator contract

Education

1. 2026 - now#### MSE Computer ScienceSUPSI, University of Applied Sciences and Arts of Southern Switzerland, Lugano Master of Engineering in Computer Science
2. 2022 - 2026#### BSc Computer Science EngineeringSUPSI, Lugano GPA 5.27 / 6.0, 180 ECTS
3. 2017 - 2022#### Scientific High School DiplomaLiceo Scientifico A. Pacinotti, La Spezia Final grade 100/100
4. #### Introduction to Metric SpacesMIT OpenCourseWare, online Self-directed study of topology and real analysis

Selected work

## Projects

How each project is labelled

- HandmadePure hand work.
- NovelNovel work, focus on new ideas not implementation.

More on this in [When I delegate, when I don't](#approach).

- ### Software Representations & Optimization Novel2026High performance computing (HPC): automatic parallelization of C programs. Speedups comparable with ad hoc parallelization, outputs verified bit-identical. 
  - Bachelor Thesis
  - High Performance Computing
  - C
  - OpenMP
- ### Machine Scheduling: Configuration LP Novel2025 - 2026Approximation algorithms for machine scheduling, with a PhD candidate and a postdoc. 
  - Semester Project
  - Linear Programming
  - Approximation Algorithms
- ### Higgs Boson Signal Detection HandmadeAn end-to-end machine learning pipeline for Higgs boson signal detection. 
  - Machine Learning
  - Jupyter
- ### Retina Blood Vessels Segmentation Handmade2025Segmentation of blood vessels in retina images: a patch-based CNN first, then a U-Net-like fully convolutional network in Keras, with data augmentation and evaluation on test masks. 
  - Computer Vision
  - Deep Learning
  - Keras
- ### 3D Graphics Engine HandmadeA 3D graphics engine in C++ with a stereo mode for VR, used to build a playable 3D Tower of Hanoi game where the discs are moved with a Leap Motion and stylized hands are drawn on screen. 
  - C++
  - GLM
  - FreeGLUT
  - VR
  - Leap Motion
- ### Graph Dynamic Programming HandmadeOptimization on graph structures and complexity analysis. 
  - Algorithms
- ### C Threads Library HandmadeA threads library in pure C, with creation, synchronization and scheduling. 
  - C
  - Systems Programming
- ### 4 points on Sphere HandmadeFormal proof that the probability of a sphere's center lying inside a tetrahedron formed by 4 random surface points is 1/8. 
  - Mathematics
- ### Smart Regression Handmade2024Automatic parametric selection of data to remove outliers and create an enhanced regression model.
- ### Pac-Man Handmade2022A recreation of Pac-Man arcade game in pure Python and Processing IDE, with automatic generation of the maze and different levels of difficulty based on the intelligence of the ghosts. 
  - Python
  - Processing IDE
- ### Arkanoid HandmadeA recreation of Arkanoid arcade game in pure Python and Processing IDE, with custom power ups and game modes. 
  - Python
### Start-Up Outzone Novel2023 - nowBackend development for the Outzone start-up. 
  - Co-Founder
  - Backend
- text token vector class score ### NLP school project HandmadeNatural language processing pipeline of text multi-class classification from data extraction and data cleaning to training and model evaluation. 
  - Jupyter

When I delegate, when I don't

## Upcasting and downcasting

AI should be used for mechanical things like formatting, boilerplate, and verification. Intellectual and creative work is done by the human.

The most difficult part is downcasting: you create information. When you upcast it is mechanical.

Upcasting takes a subset of the information. Downcasting finds new information and connects it, like adding nodes to a graph. Given all the information it needs, an LLM can only upcast: it is a semantic translator.