Operations research and quantitative modeling

Portrait of Jocelin Cailloux

I design optimization and decision-support systems for complex physical operations. I frame the problem with business teams, model it and ship software integrated into their tools.

Ph.D. from the University of Geneva, with 7 years of experience in applied optimization.
Eligible for a G permit (cross-border commuter, EU).

  • Industry: over 30,000 physical assets simulated in seconds instead of hours, planning time cut from 4 hours to 30 minutes at DIXI Polytool
  • Software: advanced C++, author of two open-source libraries, including relais, in production behind a website serving thousands of pages a day
  • Research: thesis accepted without revisions, public defense rated Excellent, article among the GSEM's best publications of the year

Deployed at DIXI Polytool, with the University of Geneva 2020–2022

Production planning optimization for precision micro‑tooling

I designed a robust optimization algorithm that groups production orders into batches to reduce stockout risk and production costs. It handles the most urgent orders first, uses as few batches as possible and builds them so as to minimize the setup cost of the worst possible schedule. I formalized the business rules with the planning team, then integrated the tool into the company's systems with its IT team. Planning time dropped from 4 hours to 30 minutes. Published in Discrete Applied Mathematics.

Recognized by the GSEM as one of the best publications of the year 

Technical details

Modeled as a graph coloring variant. Tabu search and constructive heuristics, validated against CPLEX on instances generated from the company's real data.

An interactive demo covers the classical batching and scheduling problem, of which DIXI's is a robust variant. The problem is solved live by a solver I wrote in C++.

Try the demo

Client contract with the University of Geneva 2022–2024

Maintenance simulator for a physical asset manager

I designed and built a simulator of the aging of a fleet of physical assets: failure and safety risk, replacement timing, dependencies between assets. I formalized the business rules with the client's teams, then integrated the tool into their platform. The client's previous simulator took 3 hours for 3,000 assets and applied only part of the rules. Mine, which replaced it, simulates the whole fleet, over 30,000 assets, in seconds with every rule. The client showcased it to its own customers until the contract ended.

Technical details

Year-by-year simulation of each asset's aging and wear, driven by weather, usage rate and the past inspections that update its assessed condition. Each type of maintenance operation has its own triggering rules. Planned operations must stay within financial, material and labor budgets. When a budget allows, an operation can be brought forward to make better use of it.

Each year, the computation is limited to the assets involved in an operation. The rules are fully configurable, and the code, written in C, ships as a Docker container.

Carried out at FernUniversität, on a topic proposed by a group of companies 2018–2020

Scheduling heuristics for semiconductor manufacturing

I designed and developed heuristics for a real-world scheduling problem, as part of the IDEV40 European R&D project (Horizon 2020). They limit delays, weighted by the importance of each task. The difficulty: switching from one task to the next does not take the same time in both directions, notably because of heating and cooling times, or chemical interactions. I presented this work at the MISTA 2019 international conference (Ningbo, China).

Technical details

Minimizing total weighted tardiness on multiple machines, with asymmetric sequence-dependent setup times. Constructive and improvement heuristics, written in C++.

relais

C++23 PostgreSQL

I designed and maintain an open-source C++ library that lets PostgreSQL-backed applications respond faster while putting less load on their database, which helps reduce infrastructure costs for high-traffic applications. Data already in memory is served straight from the cache, the rest is read from the database, without the application having to handle it. Concurrent queries are grouped together, which reduces the number of round trips to the database. It is used in production behind a community website serving thousands of pages a day.

View on GitHub
Technical details

Each database table is declared in C++ and configured at compile time with templates and concepts. With this design, unused features have no runtime cost, and many usage errors are caught at compile time. For high throughput, cache reads are lock-free and database queries are handled by coroutines without switching threads.

Concurrent reads of the same table are merged into a single SQL query and writes are pipelined. These optimizations also work for tables configured without the cache. The database is the source of truth and, on one instance, relais guarantees that the cache stays in sync with it, even when a network error interrupts a write. The tests check these guarantees against a real PostgreSQL database, including by cutting connections or blocking responses. With several instances, the guarantee holds when the data is partitioned, and a time to live (TTL) bounds the gap otherwise.

Optionally, the cache stays within a memory budget. Once it is reached, GDSF eviction removes the lowest-scoring entries first:

score = f×c m
f
recent access frequency of the entry
c
average cost of reloading it from the database
m
memory taken up by the entry

periple

C++20

I designed and maintain an open-source C++ library for the travelling salesman problem: finding the cheapest order of visits, at the heart of many industrial problems such as delivery routes or production sequences. The need arose from an industrial production scheduling problem, for which I had to put together six solving methods myself, chosen according to the size of the data. I am making it as general as possible, to reuse it in as many projects as I can. It returns the optimal order, with proof, or the best solution found within the time allowed. It is already at work in this site's demo, where it orders the families.

View on GitHub
Technical details

Two exact methods prove the optimum: Held‑Karp dynamic programming and a branch and bound with Lagrangian bounds. For larger problems, a constructive heuristic, nearest neighbor, followed by 2‑opt local search quickly yields an approximate solution. The library is designed so that all its methods solve every supported variant of the problem (currently time windows, with or without service times), with symmetric or asymmetric distances. The first algorithms on offer were selected so that the library would be usable in production from the start. Like them, future, more powerful algorithms will be tested automatically under every variant and compared with the known optimum of about thirty benchmark problems (TSPLIB).

Optimization

Combinatorial optimization, linear and mixed-integer programming (MILP), robust optimization (min-max), constraint programming, metaheuristics (tabu search, genetic algorithms, simulated annealing), multiobjective optimization, simulation

Programming

C++ — advanced C Python Julia SQL

C++20/23 (templates and concepts, coroutines, multithreading, atomics, memory management), Python with NumPy, Pandas and SciPy

Solvers and tools

Gurobi, CPLEX, CP Optimizer, JuMP, MiniZinc, PostgreSQL, CMake, Docker, Git, Linux, CI/CD, REST APIs

Languages

French — native
English — fluent (C1)
German — intermediate (B1)

Professional experience

  1. Ph.D. Researcher in Operations Research

    2020–2025

    University of Geneva — GSEM

    Geneva, Switzerland

    Optimization and simulation: 2 solutions deployed at industrial clients

    Research: 1 published article and 2 submitted, including a best-practice guide to framing and optimizing industrial problems with prioritized objectives

    Teaching: 5 courses, from Bachelor to MBA level

  2. Research Associate in Optimization

    2018–2020

    FernUniversität — Faculty of Mathematics and Computer Science

    Hagen, Germany

    Optimization: scheduling heuristics for semiconductor manufacturing, as part of the IDEV40 European R&D project

    Teaching: in German, remote student support and exam marking

Education

  1. Ph.D. in Management (thesis in Operations Research)

    2020–2026

    University of Geneva — GSEM

    Geneva, Switzerland

    Thesis by publication: Metaheuristics and Optimization Frameworks for Industrial Planning and Scheduling under Resource and Computation‑Time Constraints

    Advisor: Prof. Nicolas Zufferey

    Awarded on November 3, 2026. Thesis accepted without revisions, public defense rated Excellent, the highest possible rating.

    President of GAME, the GSEM doctoral students' association (2021–2023)

    Thesis abstract

    This PhD thesis investigates the design of optimization methods for complex industrial decision-making under multiple hierarchical objectives and strict computation-time constraints. It focuses on real-world production planning and scheduling problems with resource constraints. A first contribution proposes an integrated lexicographic optimization approach, combining graph-coloring and traveling-salesman formulations with efficient metaheuristics to address an industrial scheduling problem arising from a Swiss micro‑machining company. A second contribution introduces Event‑Driven Tabu Search (EDTS), an adaptive metaheuristic framework that extends the standard Tabu Search paradigm by dynamically balancing intensification and diversification based on search-state indicators, thereby reducing parameter sensitivity and improving solution quality on large-scale industrial instances. Finally, the thesis formalizes a general methodology for designing lexicographic optimization algorithms that explicitly integrates objective priorities with industrial computation-time requirements. Drawing on applications in production, supply chain management, and transportation, it identifies common design principles and provides practical guidelines for developing optimization methods that are both computationally efficient and directly applicable to industrial decision support.

  2. MSc in Operations Research

    2016–2018

    Paul Sabatier University, jointly accredited with ENAC, ISAE-SUPAERO and ENSEEIHT (M2), and University of Nantes (M1)

    Toulouse and Nantes, France

    Master's thesis: Decomposition and hybrid methods for complex scheduling problems

    Topic: batching and scheduling in semiconductor manufacturing, using constraint programming and MILP in C++

    M2 research internship: LAAS‑CNRS, supervised by Dr. Christian Artigues and Dr. Pierre Lopez

    M1 research project: exact biobjective optimization, supervised by Dr. Anthony Przybylski

  3. BSc in Computer Science

    2013–2016

    University of Nantes

    Nantes, France

Discrete Applied Mathematics 2024

Graph coloring approaches for a production planning problem with makespan and setup penalties in a product-wheel context

Cailloux J., Zufferey N., Gallay O.

The methods behind the tool deployed at DIXI Polytool: a tabu search, limited to the 30 minutes of computing time set by the company. On 216 instances generated from its data, with up to 200 production orders, it cuts the setup time of the worst possible schedule by more than 20% on average compared with the shop rule. CPLEX, given 2 hours per objective, no longer finds a complete solution from 100 orders on.

Recognized by the GSEM as one of the best publications of the year 

Read article 
Abstract

In this paper, we introduce a clustering and scheduling problem on a production line modeled as a single machine. A set of jobs (some of them being urgent) must be partitioned into clusters, and a robust (with respect to a min–max criterion) cyclic sequencing of the clusters must be determined (i.e., the product-wheel paradigm is employed). Each cluster has to satisfy two constraints: the setup constraint (i.e., only jobs with small setup times between them are allowed in the cluster) and the capacity constraint (i.e., the setup and processing times in the cluster cannot exceed a given shift duration). Three objective functions are minimized in a lexicographic fashion: (1) the number of urgent clusters (i.e., containing at least one urgent job); (2) the total number of clusters; (3) a worst-case scenario with respect to the setup among clusters. In other words, makespan and setup penalties are considered. Graph-coloring models and methods are designed for (1) and (2), whereas traveling-salesman approaches are introduced for (3). The considered problem was proposed by a micro-machining company located in Switzerland, named DIXI polytool. In order to cover their industrial needs, the company imposed very strict computing-time limitations (a few minutes only), and was able to provide realistic instances with different characteristics. Three methods are compared in our experiments: an integer linear model (with CPLEX), a constructive heuristic that represents a current-practice rule, and a metaheuristic relying on various tabu-search procedures. Results show the efficiency (with respect to quality and speed) of our metaheuristic, and managerial insights are provided.

ODS 2022 (Florence, Italy), ROADEF 2025 (Champs-sur-Marne, France)

European Journal of Operational Research

Event-Driven Tabu Search and its Application to an Industrial Production Scheduling Problem

Cailloux J., Zufferey N., Lurkin V., Ghaddar B.

A tabu search method designed for a wide range of industrial problems, which adjusts its own search parameters based on the solutions it finds. Tested on a real production scheduling problem, it finds on every run 21 of the 22 optima proven by CPLEX, and gives the best solution of all the methods compared on 60 of the 63 instances, with up to 300 production orders. With 200 and 300 orders, standard tabu search and the shop rule exceed the setup time of that best solution by more than 50% on average.

Abstract

We introduce Event-Driven Tabu Search (EDTS), an adaptive single-solution metaheuristic that extends classical tabu search through an event-driven control strategy. EDTS monitors search indicators capturing solution quality, diversity, originality, and stagnation. Event-response rules dynamically regulate neighborhood sampling, improvement strategies, restart amplitude, and memory-based diversification. We apply EDTS to an industrial production scheduling problem that combines clustering and cyclic sequencing decisions, where each candidate clustering is evaluated by solving an embedded Traveling Salesman Problem. Computational experiments on 63 benchmark instances reflecting industrial operating conditions show that EDTS consistently outperforms an industrial constructive heuristic, standard tabu search, and tabu search with restarts. On the largest instances with up to 300 jobs, EDTS maintains an average gap below 10% relative to the best-known solutions, while the competing heuristics exceed 50%. These results demonstrate the potential of event-driven adaptation to improve the scalability and robustness of tabu search for computationally expensive scheduling problems.

RAIRO - Operations Research

A Framework for Designing Time‑Constrained Lexicographic Optimization Methods for Industrial Decision-Making

Zufferey N., Janjevic M., Cailloux J.

A methodology and good practices for designing efficient optimization methods, metaheuristic or exact, within the computing time set by the company. In a multi-objective context, the central practice is ranking the objectives by priority, which allows solutions to be evaluated immediately and the problem to be decomposed into subproblems, one per objective. To handle uncertainty, simulation evaluates only the most promising solutions when it is costly, and alternates with reoptimization when it is fast. The paper draws on five industrial projects in production, supply chain and transportation, with computing times from 1 second to 30 minutes.

Abstract

Multi-objective optimization is a key tool for supporting industrial decision-making, where organizations must balance multiple, often conflicting objectives. Traditional multi-objective methods often rely on subjective weighting, which is difficult to define and justify in real-world contexts, and typically overlook computation time as a critical design constraint. Lexicographic optimization, despite its practical relevance, remains underexplored in industrial applications. This paper proposes a formal methodology for designing lexicographic-optimization approaches that explicitly integrate decision-makers' computation-time limits and objective hierarchies. The methodology is illustrated through multiple real-world industrial cases in production, supply chain management, and transportation, from which we extract generalizable features of effective lexicographic methods. Finally, we provide practical guidelines for designing lexicographic algorithms tailored to industrial problems.

Teaching and research assistant 2020–2025

University of Geneva — GSEM

5 courses and 300 students a year, from Bachelor to MBA level, in French and English. Exercise sessions, a yearly theory lecture and, at Master level, supervision and grading of individual projects: identifying a business optimization problem, modeling it as a MILP, solving it in Excel, then running a sensitivity analysis.

  • Optimisation et gestion des opérations Bachelor · 5 years
  • Gestion de la production et des stocks Bachelor · 4 years
  • Prescriptive analytics Master · 5 years
  • Operations, logistics and supply chain management Master · 5 years
  • Organisation de la production MBA, working professionals · 1 year

Research Associate in Optimization 2018–2020

FernUniversität — Faculty of Mathematics and Computer Science

Distance teaching at Master level, in German: answering students' questions and marking exams.

  • Entscheidungsmethoden in unternehmensweiten Softwaresystemen Decision methods in enterprise-wide software systems