Polytech Lille · Differential Algebra & HPC & AI · Seeking PhD

I'm an engineering student at Polytech Lille (graduating August 2027). During my 3rd-year internship at CRIStAL (UMR 9189, Université de Lille), I worked under the supervision of François Boulier on differential algebra and the algorithmic resolution of algebraic differential equations — specifically, detecting Puiseux series solutions via monomial changes of variable.

I'm currently completing my 4th-year internship at NTT Data, where I work on advanced optimization of truck fleet problems using machine learning, deep learning, reinforcement learning, and Automated Heuristic Design (AHD) with metaheuristics. I also contribute to quantum security projects involving Quantum Key Distribution (QKD) and Post-Quantum Cryptography (PQC).

Starting September 2026, I'll be on an Erasmus exchange at Politecnico di Milano (Polimi) for my final semester.

I build projects at the intersection of mathematical computing, high-performance computing, and artificial intelligence. My work spans from custom CUDA kernels for transformer training to bit-level optimized Monte Carlo simulations. I'm seeking a PhD starting September 2027 in AI, HPC, or applied mathematics.

Developed a complete method for detecting and computing Puiseux series solutions (fractional-exponent power series) of algebraic ordinary differential equations, via the monomial change of variable $x^a = t^b$.

Key results:

• Congruence condition for efficient FPS existence filtering before costly differential elimination
• Completeness theorem: monomial changes suffice among all algebraic changes of variable
• Effective bound $b \leq \deg(P)$ via Newton differential polygon
• Two algorithms: fast_decision ($O(N \cdot D^2)$ without Denef–Lipshitz) and full_solver

→ puiseux-differential-algebra — Code, algorithms, and 15-page paper.

A GPT language model built entirely from scratch in C++ and CUDA — no PyTorch, no TensorFlow. Custom kernels for Softmax, GELU, RMSNorm, AdamW, and cross-entropy. Uses cuBLAS for matrix multiplies.

• ~2.7M parameters, 6 transformer blocks, 192-dim embeddings
• Trained on Shakespeare (1.1M chars) on NVIDIA T4 — 5.51 → 1.19 loss
• Xavier init, AdamW, learnable RMSNorm, LR warmup + cosine decay
• All forward/backward passes hand-written

→ Transformer-Cuda-Framework

A custom chess engine in C with bitboard move generation, NNUE evaluation (768→256→32→1), and alpha-beta search with Null Move Pruning and Late Move Reductions — plus a full reinforcement learning pipeline (self-play in C++ with LibTorch → Policy Gradient in Python).

51.8 million nodes/sec in pure move generation. Search speed is limited by the missing transposition table — a planned future addition.

→ NeuralBitChess

Real-time 3D Ising model simulation and visualization with PyVista. Manipulate temperature and magnetic field in real-time and observe phase transitions at 60+ FPS.

• Bit-packing: 64 spins per uint64_t — 64M total spins
• Software binary adder: computes all 64 spin energies simultaneously via bitwise operations
• Checkerboard decomposition: deterministic red-black parallel updates
• 128M spins/sec on a single M1 core (11× over standard Metropolis)

→ Ising-3D-Optimized

Numerical study of the phase transition in the 2D Ising model using the Metropolis-Hastings algorithm, implemented across three HPC approaches: single-threaded C (30M steps/sec), multi-threaded with pthreads (33M steps/sec), and CUDA GPU (~2B steps/sec). Validates Onsager's exact solution ($T_c = 2.269$).

→ Ising-MCMC-HPC · Grade: 19/20

• puiseux-differential-algebra — Puiseux solutions of AODEs via $x^a = t^b$ · CRIStAL

• Transformer-Cuda-Framework — GPT from scratch in C++/CUDA
• NeuralBitChess — Chess engine with NNUE + RL pipeline
• Ising-3D-Optimized — Real-time 3D Ising with bit-parallel HPC
• Ising-MCMC-HPC — 2D Ising: CPU vs pthreads vs CUDA · 19/20
• Derivative-LLM — Transformer trained to differentiate polynomials
• Derivative-Agent — LLM-powered agent for math operations

• PageRank — PageRank on real navigation data · 20/20
• cartPole — Deep Q-Network on custom environment
• financial-stats — Moving average strategy analyzer · Live Demo

• PhotoDetector — macOS desktop app for YOLOv8 object detection
• Projet-Graphe — Maximum flow (Dinic algorithm) · 19.5/20
• school-dashboard — Gender salary gap visualization · 17.5/20
• Personal-Neovim — Modular Neovim configuration
• ubersicht-widget-crypto — Real-time crypto prices on macOS desktop

• Portfolio: alexgit31.github.io/Portfolio
• LinkedIn: linkedin.com/in/alexis-evaristo
• Email: alexis.evaristo@polytech-lille.fr
• Location: Lille, France

Polytech Lille (graduating Aug 2027) · Erasmus @ Polimi (Sep 2026–Feb 2027)
Seeking a PhD starting September 2027 in AI, HPC, or applied mathematics.