William Stoops: I choose an architecture, I measure it, I ship it.

Journey

1. 2021 · 1st year

   Epitech

   I join Epitech for a Master of Science, Expert in Information Technology: class of 2026.

2. 2022 · 2nd year

   GDS Élec

   Full Stack Engineer, GDS Élec

   Jul 2022 – Jan 2023

   • I shipped an application to production that manages electric vehicle charging stations remotely, over the OCPP protocol.

3. 2023 · 3rd year

   Launching STAXX, Strattt, then INTM

   I launch STAXX, my final-year project, which I carry through to my 5th year.

   Full Stack Engineer, Strattt

   Sep 2023 – Feb 2024

   • I automated an accounting pipeline end to end.

   Full Stack Engineer, INTM Groupe

   2024

   IT services and consulting company.

   • I delivered alone and from scratch the company’s internal activity management tool: business managers’ KPIs, consultant status tracking (in training, on assignment, at which client).

   • NestJS
   • React
   • PostgreSQL

   I am happy to recommend William Stoops, who did an internship at INTM Group. I had the pleasure of working with him on React development assignments, and although I officially had a mentoring role, it was clear from the start that William already had a solid technical background and great autonomy.

   William stood out for his curiosity and his will to learn. He approached every challenge with enthusiasm and always sought to deepen his knowledge and skills. His autonomy and motivation were obvious, making our collaboration both productive and rewarding.

   I am convinced that William will keep excelling in his career, and I warmly recommend him to any team looking for a talented and motivated developer. His drive and thirst for learning are major assets for any project.

4. 2024 · 4th year

   South Korea, Korea University

   고려대학교 — 한국 — 딥러닝 — 컴퓨터 비전 — 고려대학교 — 한국 — 딥러닝 — 컴퓨터 비전 —

   A year at Korea University, in South Korea, taught in English: deep learning and computer vision.

   France한국South Korea

   안녕하세요안녕하세요(hello)

   France한국South Korea

   한국 · South Korea

   고려대학교

   Korea University

   • 61st in the QS World University Rankings
   • 1 year of courses in English
   • 3 models trained

   What I learned there

   A year to go from using models to building them: the mathematics first, then how a network learns, computation by computation, and the architectures that came of it.

   The mathematics

   Linear algebra

   On Gilbert Strang’s textbook, Introduction to Linear Algebra

   • Matrices and factorizations
   • Vector spaces
   • Orthogonality and least squares
   • Eigenvalues
   • SVD
   • Backpropagation and stochastic gradient descent

   Learning

   Forward and backward propagation

   Computed step by step: the output layer by layer, then the error’s gradient carried back by the chain rule.

   Supervised learning

   Learning from labelled examples, narrowing the gap between prediction and truth.

   Reinforcement learning

   Learning by acting: an agent, an environment, a reward to maximize.

   The architectures

   1. MLPDense layers: the basic network.
   2. CNNConvolutions, for images.
   3. RNNA memory, for sequences.
   4. LSTMA long memory, keeping what matters.
   5. TransformerAttention, at the heart of LLMs.

   Models trained

   • Fine-tuned BERT

     Text classification, with Hugging Face.

   • CNN

     The state of a chess game, recognized from an image of the board.

   • Gesture detector

     In real time, YOLO-style, on a webcam feed.

   

   

5. 2025 · 5th year

   Back in France

   안녕히 계세요(goodbye)

   France한국South Korea

   Bonjour.Bonjour.(hello)

   한국South KoreaFrance

   Back in France, I join IT-Finance, publisher of ProRealTime. With the STAXX team, we go on NRJ Lille, then win the Epitech Summit.

   Software Engineer, IT-Finance, publisher of ProRealTime

   Since Sep 2025

   Financial software publisher, about 70 people.

   • I designed and shipped to production, in C++, an implied volatility computation the product did not have, across the OPRA options universe, from studying the models to deployment. Sole developer on the subject. Its results now serve tens of thousands of options traders.
   • The computation runs continuously and had drifted to ten hours per cycle. Against the team’s hypothesis, which blamed the algorithm, I proved by measurement that the cost came from the data structure. I redesigned it: from 10 hours to 5 minutes, and values up to date in the product again.
   • I migrated the platform’s financial news service, in place for years, to Rust, and added a cache: 99% lower latency on most requests. In production, serving hundreds of thousands of users. A language learned on the job.

   • C++
   • Rust
   • Python

   Three-dimensional implied volatility surface, by strike and maturity. It forms a valley: volatility is lowest around the current price and rises on distant strikes, most at short maturities; over time, the curve flattens and rises slightly.

   Turn the surface

   Turn leftTurn rightTilt upTilt down

   The computation, in miniature

   Implied volatility

   An option’s price depends on the volatility expected of the underlying. Implied volatility is the one that, in the Black-Scholes model, gives back the observed price: one per strike and maturity. Together, they form this surface.

   σ 15%30%

   Drag or use the arrow keys to turn. Hover to read the smile and the term structure.

   One full cycle, to scale

   1 hour = 1.2 seconds

   Before the redesign10 h per cycle

   0 / 1 cycle

   After5 min per cycle

   0 cycles

   Run both computations

   Simulated time: 0 h 00

   While one old cycle completes, the redesigned version runs 120: values up to date again.

   Since 2024

   STAXX

   Equipment ordering platform for the construction industry

   • 1st at the Epitech Summit competition
   • 300 people at the pitch
   • 3 developers, two of whom I led

   Epitech final-year project. 1st at the Epitech Summit competition, pitched to 300 people.

   • I drove the project and led the two other developers: splitting the work, architecture decisions, technical consistency through to delivery. I built the entire back end.
   • I designed the matching engine between the language of site crews and the catalogue: “20 Atlas clamps for 26 mm tube” finds the right item despite typos, through an alias dictionary that every user correction enriches.

   • NestJS
   • PostgreSQL
   • React

   

   

   Play the video: STAXX pitch at the Epitech Summit competitionPlay the videoSTAXX pitch at the Epitech Summit competition

   

6. 2026 · Class of 2026

   Today

   Epitech class of 2026. Software Engineer at IT-Finance, publisher of ProRealTime.