Review of Atoptima, DeepTech Optimization Software Vendor

By Léon Levinas-Ménard

Last updated: November, 2025

Atoptima is a Bordeaux-based software editor created in 2019 as a deeptech spin-off from the RealOpt research team at Inria, CNRS and the University of Bordeaux, specializing in advanced mathematical optimization for complex planning problems in logistics, warehousing, production and network design.¹² The company builds a suite of cloud-hosted optimization solvers—RouteSolver for vehicle routing, PackSolver for 3D palletization and loading, PickSolver for warehouse slotting and picking, PlanSolver for production and workforce scheduling, and FlowSolver for network flow and consolidation—delivered as SaaS applications and as asynchronous APIs through an orchestration layer called Galia.²³⁴⁵⁶ Technically, Atoptima’s stack is centered on Julia: its open-source Coluna.jl and BlockDecomposition.jl frameworks implement branch-and-price / branch-cut-and-price algorithms for block-structured mixed-integer programs, integrated with JuMP and external MIP solvers such as HiGHS, GLPK, Gurobi and CPLEX; the commercial solvers build on these components to compute operational decisions like optimized tours, loading patterns and schedules.⁷⁸⁹¹⁰¹¹¹² Commercially, Atoptima remains a small but active vendor—roughly fifteen employees, one disclosed €1.2M seed round in 2021 plus an i-Nov innovation grant in 2024—with a handful of verifiable client references such as Logtran (serving Carrefour-group banners in the French Antilles), CMA CGM / CEVA Logistics (fleet decarbonization planning) and AppliColis (urban cycle logistics), alongside numerous anonymized case studies in logistics and industry.²¹³¹⁴¹⁵¹⁶¹⁷¹⁸¹⁹²⁰²¹²²²³²⁴²⁵²⁶ Atoptima’s technology is clearly state-of-the-art in classical mathematical optimization, but it does not appear to cover demand forecasting or probabilistic inventory optimization; instead, it offers high-end prescriptive modules that assume upstream demand and cost inputs and return optimized routing, packing and scheduling decisions.

Atoptima overview

Atoptima positions itself as a “decision-support software publisher” industrializing more than two decades of academic research in combinatorial optimization into operational tools for logistics and production planning.¹²¹² Spun out in 2019 from the RealOpt team in Bordeaux, it focuses on discrete planning problems such as vehicle routing with time windows, depot location, 3D palletization, warehouse picking, and machine or workforce scheduling rather than on broad end-to-end supply chain suites.¹²³⁴²⁷ The offer consists of a family of optimization modules—RouteSolver, PackSolver, PickSolver, PlanSolver and FlowSolver—exposed either through a SaaS web interface where users upload CSV/JSON data, run optimizations and visualize maps or Gantt charts, or through an asynchronous API layer, Galia, that accepts optimization jobs and returns results via webhooks and WebSockets.²³⁴⁵⁶²⁷ Under the hood, the company’s algorithms rely on Julia-based open-source projects such as Coluna.jl and BlockDecomposition.jl, which implement Dantzig-Wolfe and Benders decompositions and branch-cut-and-price strategies for block-structured mixed-integer programs; these are combined with commercial or open-source MIP solvers via JuMP and MathOptInterface.⁷⁸⁹¹⁰¹¹ The firm is small and deeptech-oriented, with around fifteen employees, one disclosed equity round of €1.2M in 2021 and additional public funding through Bpifrance, ADEME and the i-Nov innovation contest; it serves a limited but nontrivial set of named clients, mostly in France and francophone markets, and many anonymized customers in logistics and retail.²¹³¹⁴¹⁵¹⁶¹⁷¹⁸¹⁹²⁰²¹²²²³²⁴²⁵²⁶²⁷ The technology is advanced in the sense of exact and decomposition-based optimization, while the commercial maturity remains that of an early-stage, project-driven vendor.

Atoptima vs Lokad

Atoptima and Lokad both address supply chain planning problems but from markedly different angles. Atoptima is a prescriptive optimization specialist: it focuses on solving large NP-hard combinatorial problems such as vehicle routing, palletization and scheduling using exact or near-exact mathematical programming, implemented in Julia on top of Coluna and BlockDecomposition, and exposed as domain-specific solvers (RouteSolver, PackSolver, PickSolver, PlanSolver, FlowSolver) that assume demand and other inputs are already known.²³⁴⁷⁸⁹¹⁰¹¹²⁷ Lokad, by contrast, is a forecasting-and-optimization platform whose core deliverable is a bespoke probabilistic optimization app that unifies demand forecasting, inventory planning, production scheduling and sometimes pricing within one Envision-based pipeline; its stack is centered on .NET (F#/C#), a custom DSL, probabilistic forecasting and stochastic optimization algorithms such as Stochastic Discrete Descent and Latent Optimization rather than on generic MIP solvers.²⁸²⁹ Atoptima’s modules typically take as input concrete tasks and resources (demand already quantified, capacities given, costs specified) and return optimized plans—routes, loading patterns, schedules—that can be plugged into existing TMS/WMS/ERP systems; it does not, based on public evidence, estimate demand distributions or automatically derive inventory policies, and its “AI” messaging mainly refers to mathematical optimization.²³⁴³⁰²³²⁷ Lokad, in contrast, starts from raw transactional and master data and builds probabilistic demand models, then computes financially optimized decisions (replenishment quantities, stock allocations, repair schedules, sometimes prices) that explicitly account for uncertainty and economic drivers such as holding costs, stockout penalties and basket effects.²⁸²⁹ Architecturally, Atoptima exposes a set of microservices around Galia, designed for asynchronous job submission of optimization tasks and integration into other systems, whereas Lokad provides a multi-tenant SaaS workspace where clients run Envision scripts on a shared compute cluster and push or pull data via SFTP and APIs.³¹³⁴¹¹⁵⁶²⁹ In terms of scope, Atoptima is narrower and deeper on specific operational problems like routing and 3D loading, grounded in decades of column-generation research; Lokad is broader across the supply chain, trading some algorithmic exactness on individual combinatorial problems for scale, probabilistic modeling and end-to-end decision pipelines. Finally, commercial maturity differs: Atoptima is a small early-stage vendor with a handful of named references (Logtran, CMA CGM / CEVA, AppliColis) and projectized deployments, whereas Lokad is a more mature, bootstrapped company dating from 2008 with a wider roster of large retail, manufacturing and aerospace clients and a long-running platform evolution focused on probabilistic forecasting and quantitative supply chain.¹²¹³¹⁵¹⁶¹⁷¹⁸¹⁹²⁰²¹²²²³²⁴²⁵²⁸²⁹

Company history and funding

Founding and academic roots

Atoptima emerged in 2019 as a spin-off from the RealOpt research team, a joint CNRS/Inria/University of Bordeaux/Bordeaux INP group specializing in combinatorial optimization and mathematical programming.¹²¹² Inria describes Atoptima as a decision-support software publisher born from long-standing collaborations with industrial partners on vehicle routing and related problems, aiming to industrialize academic advances in optimization.¹ CNRS Innovation similarly presents Atoptima as the culmination of over 25 years of expertise in combinatorial optimization accumulated within RealOpt, with a mission to turn cutting-edge mathematical models into usable software tools.²¹² The founding team includes François Vanderbeck (scientific director and long-time OR professor), Vitor Nesello and Adrien Duruisseau, combining mathematical, engineering and business backgrounds.²¹² Headquartered in Bordeaux, the company deliberately positions itself as a deeptech player bridging academic research and industrial deployment rather than as a generic enterprise software vendor.¹²

Funding and public support

Public sources consistently report a single equity funding round for Atoptima: a €1.2 M seed round in 2021 led by Epopée Gestion, Bpifrance, ADEME and Région Nouvelle-Aquitaine.¹³¹⁴ Tracxn lists this as the company’s total disclosed funding, equivalent to roughly US$1.4 M, with the latest round dated September 2021.¹³ Societe.Tech and several French business press articles (notably coverage of the funding round) corroborate the amount and investor line-up, framing the funding as a means to accelerate product industrialization and international deployment.¹⁴ In addition to equity, Atoptima obtained public R&D support through the i-Nov innovation contest: CNRS Innovation reports that in 2024 Atoptima won i-Nov with a project budget of €1,113,177, of which €500,929 was subsidized, aimed at enhancing its optimization solutions and market reach.² The company has also benefited from French Tech selections (e.g. French Tech NA20) and from incubation support by Unitec in Bordeaux.¹² Revenue estimates from Compworth (around US$870k) should be treated as indicative rather than precise, but they are consistent with a small but commercially active deeptech company.¹²

Acquisition activity

Searches of CB Insights, Tracxn and other startup databases show no acquisitions involving Atoptima—neither as acquirer nor as target.¹³³⁰¹⁵ Press coverage, CNRS/Inria profiles and company materials likewise make no reference to mergers or acquisitions, focusing instead on organic development and public funding.¹² On available evidence, Atoptima remains an independent startup with no M&A history as of late 2025.

Product portfolio and supply chain scope

Atoptima’s product line is best described as a suite of domain-specific optimization solvers for discrete planning problems in operations and supply chain.²³⁴²⁷ Across CNRS Innovation, FAQ Logistique and Atoptima’s own site, the same core modules recur:

RouteSolver – vehicle routing and transportation optimization (multi-depot VRP, pickup & delivery, time windows, multi-period routing, multimodal routes, fleet sizing and tour planning).²³⁴²⁷
FlowSolver – optimization of flows across a logistics network, including consolidation at hubs, cross-docking, and multi-echelon routing.²³⁴
PackSolver – 3D palletization and loading (truck, container or ULD loading, pallet building, packing oversized items) with volumetric and stability constraints.²³⁴²⁷
PickSolver – warehouse slotting and order picking, including SKU-to-location assignment, batching and picker route structuring.²³⁴²⁷
PlanSolver – production and workforce scheduling, including sequencing on machines, lot-sizing, shift planning and timetabling.²³⁴

These modules are typically embedded into or complement existing systems rather than replacing them. FAQ Logistique explicitly frames Atoptima’s offer as a way to enrich TMS, DMS, WMS, OMS or APS with advanced decision-support, rather than as a full planning suite.²⁷ CNRS Innovation stresses that the tools are used at multiple planning levels—strategic, tactical and operational—for problems such as depot location, transport planning, production scheduling and workforce planning.² Notably, there is no evidence of built-in demand forecasting or inventory policy optimization; Atoptima’s modules assume that the demand, cost and constraint parameters are provided and focus on computing prescriptive decisions (routes, packing patterns, schedules, resource allocations) given those inputs.²³⁴²⁷

Technology stack and architecture

Optimization engine: Coluna and BlockDecomposition

Atoptima’s most distinctive technical asset is its open-source optimization stack centered on Coluna.jl and BlockDecomposition.jl, developed in collaboration with academic partners.⁷⁸⁹¹⁰¹¹¹² Coluna is a branch-and-price-and-cut framework implemented in Julia: users model mixed-integer problems in JuMP, annotate their block structure using BlockDecomposition, and let Coluna reformulate them via Dantzig-Wolfe or Benders decomposition before applying branch-cut-and-price algorithms.⁷⁸⁹¹⁰ BlockDecomposition extends JuMP with macros that declare master and subproblems, define axis sets and specify how variables and constraints are grouped, enabling generic implementation of decomposition schemes.⁸⁹

Coluna integrates with multiple LP/MIP solvers through MathOptInterface (HiGHS, GLPK, Gurobi, CPLEX, among others), allowing Atoptima to combine decomposition with the strengths of existing solvers.⁷⁸⁵ Presentations at events like the Column Generation conference and MINOA workshops emphasize its focus on block-structured MILPs typical of logistics and industrial applications—vehicle routing, cutting stock, location-routing, scheduling—where standard flat formulations are either too slow or too large.⁹¹⁰¹² The RealOpt publication record reinforces this picture: decades of work on column generation, primal heuristics, diving strategies and stabilization techniques for practical routing and cutting problems feed directly into Coluna’s design.¹²

Although Atoptima does not explicitly state that the commercial RouteSolver, PackSolver, PickSolver, PlanSolver and FlowSolver are built on Coluna, the overlap of team, technology and problem classes makes it highly plausible that these solvers are essentially domain-packaged layers on top of Coluna/BlockDecomposition, tailored for specific industries and integrated into the SaaS platform.²³⁴⁷⁸⁹¹⁰¹¹¹²

System and integration architecture

Job postings and GitHub repositories reveal the broader system architecture: Atoptima’s “Apps are developed in Julia and cloud-based as micro-services,” with in-house software development and open-source collaboration with academic labs.³¹³² The company’s GitHub organization lists multiple Julia libraries (Coluna.jl, BlockDecomposition.jl, DynamicSparseArrays.jl, forks of Redis.jl and JSON3.jl) alongside a minimal Galia JavaScript client, indicating a service-oriented back-end and JavaScript-based integration layer.⁷⁸¹¹⁵

The Galia platform is central to integration. The minimal-galia-js-client repository demonstrates how to submit optimization jobs to Galia and receive results asynchronously via webhooks and WebSockets, using environment variables such as GALIA_HOST, APPLICATION_ID and ACCESS_TOKEN.⁵ The galia.atoptima.com domain exposes a “GaliaFrontEnd” login, suggesting a web interface to monitor or manage jobs.⁶ From these artifacts, one can infer an architecture where:

Julia microservices implement each solver, interacting with Coluna and external MIP solvers.
Redis and related components support caching, job queues or state management (as implied by the Redis.jl fork).¹¹
Galia orchestrates job submission, queuing and result delivery, decoupling compute-intensive optimization runs from client applications.⁵⁶
Web applications (likely single-page apps) provide CSV/JSON upload, visualization (maps, Gantt charts, lists) and scenario management for human planners, while external systems can integrate via the Galia API.²³¹³⁴⁵⁶

Atoptima’s “Solutions / How it works” pages describe the user-level workflow: upload tasks and resources as data, inspect and adjust them, launch optimization, explore results on maps or timelines, tweak scenarios (e.g., manual adjustments), and export solutions.³⁴ The architecture is clearly more than CRUD; the complexity lies in the optimization back-end, while the UI and API layers are relatively conventional web technologies.

Deployment and roll-out

CNRS Innovation and Atoptima’s case materials describe a project-based deployment model with relatively short implementation cycles.²⁶ Solutions are delivered in SaaS mode, with two primary access modes:

A simplified web application for quicker onboarding and standard problem classes.
A more parameterizable module (via Galia and deeper configuration) for tailored, complex use cases.²³⁴⁵

CNRS indicates typical deployment times of two to six weeks, including problem scoping, data integration, solver configuration and validation.² In a multi-solver case study for a multinational in transport equipment, Atoptima reports a three-week “solver setup time” for a solution combining RouteSolver, PlanSolver, PackSolver and FlowSolver to optimize inbound/outbound flows, multimodal routing and 3D loading.⁶

Inria and FAQ Logistique emphasize that Atoptima’s modules are meant to integrate with existing TMS/WMS/ERP systems rather than replace them, and that the integration footprint is relatively light thanks to the SaaS and API model.¹²⁷ However, detailed information on hosting providers, SLAs, multi-tenancy, data residency and security certifications (ISO 27001, SOC2, etc.) is not publicly available, leaving non-functional characteristics largely undocumented.

Clients, sectors and geographies

Named, verifiable references

The Logtran deployment is Atoptima’s most documented production case. Atoptima’s blog describes how Logtran, logistics and transport provider in the French Antilles & Guyane (part of the Safo group), adopted Atoptima’s software to optimize palletization, truck loading and distribution routes, targeting and achieving around 20% transport cost reduction.¹⁶¹⁷ Supply Chain Magazine confirms that Logtran chose Atoptima’s SaaS solution to optimize tours and truck loading for deliveries in French overseas territories.¹⁸ Voxlog similarly reports a 20% cost reduction and highlights the use of Atoptima’s optimization for both tours and loading.¹⁹ Stratégies Logistique adds that Logtran serves Carrefour, Proxi, 8 à Huit and Promocash stores in the French Antilles, making Atoptima’s engine part of the logistics backbone supplying these retail banners.²⁰

In the Smart Port Challenge context, Atoptima was selected by CMA CGM to co-develop a decision-support tool for decarbonizing road transport by planning the deployment and allocation of zero-emission heavy trucks.²¹ SITL Daily and CCI Marseille-Provence both report that Atoptima was chosen as a laureate to work with CMA CGM on tools to accelerate the transition to electric and hydrogen trucks.²¹²² Atoptima’s own “decision-making AI” blog further explains that it collaborated with CMA CGM and CEVA Logistics on strategic positioning of new charging stations and tactical allocation of zero-emission vehicles across warehouses and flows.²³

Atoptima has also co-developed CycloCo, an urban last-mile logistics platform, with AppliColis. A joint press release (AppliColis–Atoptima) describes an ADEME-supported project to create a centralized system for multimodal, sustainable last-mile logistics, with Atoptima contributing the optimization software.²⁴ An Atoptima blog on “AI and green supply chain” discusses CycloCo as a centralized system for ecological urban delivery, while an ADEME publication documents a project titled “Plateforme de planification dynamique pour la cyclologistique urbaine,” co-authored by AppliColis and Atoptima.²⁵²⁶

During the COVID-19 crisis, FAQ Logistique reports that Atoptima made its expertise available to healthcare actors by deploying vehicle routing optimization applications for ambulance and patient transport between hospitals, free of charge, to improve the efficiency of emergency logistics.³³

Anonymized cases and sector coverage

Atoptima’s website and FAQ Logistique list multiple anonymized or vaguely described clients, such as a “multinational in transport equipment,” a “world leader in logistics,” a “European leader in mass retail,” and actors in express delivery and maintenance routing.²⁶²⁷ These customer stories detail problem types (complex inbound/outbound flows, high service-level needs, CO₂ reduction targets) and performance claims (e.g. 30% cost savings, 20% CO₂ reduction, 30% productivity gains) but do not name the companies involved, making independent verification impossible.²²⁷

The verifiable named clients and projects place Atoptima primarily in France and francophone markets (Metropolitan France, French Antilles & Guyane) with activity in:

Transport and logistics (Logtran, CMA CGM/CEVA, unnamed LSPs).¹⁶¹⁷¹⁸¹⁹²⁰²¹²²²³
Retail grocery distribution (via Logtran’s servicing of Carrefour-group banners).²⁰
Urban last-mile and cycle logistics (AppliColis / CycloCo).²⁴²⁵²⁶
Healthcare logistics (ambulance routing projects during COVID-19).³³

Given the limited number of named references, Atoptima appears to have early but concrete traction in these segments, with additional anonymized projects pointing to a somewhat broader but less verifiable client base.

Assessment of technical claims

“AI” and machine learning

Atoptima frequently uses the language of “AI” and “decision-making intelligence”, but the technical artifacts point overwhelmingly toward deterministic mathematical optimization, not machine learning. Its scientific background, as presented by CNRS, Inria and RealOpt, is centered on column generation, branch-and-price, cutting stock, routing, scheduling and related OR techniques; there are no public publications or code artifacts indicating production use of regression models, deep learning or reinforcement learning.¹²⁷⁸⁹¹⁰¹¹¹² The open-source stack (Coluna, BlockDecomposition, DynamicSparseArrays) implements decomposition-based MILP algorithms and low-level numerical data structures, not ML infrastructure.⁷⁸¹¹

FAQ Logistique and CNRS do describe Atoptima’s tools as “decision AI” or “artificial intelligence software,” but the examples they give—route optimization, 3D loading, network design, scheduling—cluster around optimization tasks.²²⁷ Atoptima’s own blog posts on “decision-making AI” and “AI for a green supply chain” frame AI in terms of automated optimization of transport networks and deployment plans, rather than predictive modeling.²³²⁵ On available evidence, Atoptima’s “AI” label essentially denotes advanced OR-based decision engines rather than machine learning systems. Any interpretation beyond that (e.g., ML-based demand forecasting or learning-based heuristics) would be speculative.

Performance, scalability and robustness

CNRS Innovation and FAQ Logistique report performance claims such as 30% cost reduction, 20% CO₂ reduction, and solvers “40 times faster than market tools,” along with productivity improvements of around 30% in certain use-cases.²²⁷ Case material on Logtran and anonymized clients echoes substantial gains in distance traveled, cost and utilization.¹⁶¹⁷¹⁸¹⁹²⁰ However, these figures are derived from vendor-provided case studies and trade-press articles rather than from independently benchmarked or peer-reviewed comparisons. No public standardized benchmarks against alternative commercial solvers or open-source libraries (e.g. VRP solvers, packing heuristics) are available, and methodological details (test sets, baselines, hardware) are generally absent.

On the other hand, Coluna’s design and the RealOpt research record strongly suggest that Atoptima’s engines are capable of tackling large-scale real-world instances that are intractable with naive MILP formulations, especially in routing and cutting stock.⁷⁸⁹¹⁰¹² The use of decomposition, dynamic column generation and stabilization is state-of-the-art practice for these problem types, and Coluna’s integration with multiple high-performance MIP solvers further supports scalability claims at the engine level.⁷⁸⁹¹⁰ What remains unclear is how much of this potential is realized in day-to-day commercial deployments, where time limits, hardware constraints and changes in data quality may necessitate heuristic shortcuts.

Robustness with respect to uncertainty is also an open question. Inria notes that handling multi-level decisions and uncertainties remains an ongoing research topic for Atoptima and its academic partners.¹ There is no publicly documented framework for stochastic or robust optimization beyond what can be encoded in deterministic models (e.g., safety margins baked into constraints). This stands in contrast to vendors that explicitly model demand and lead time uncertainty; Atoptima’s strength lies in deterministic combinatorial optimization, with uncertainty handled, if at all, outside the solver.

Commercial maturity

Bringing together headcount, funding, references and deployment stories, Atoptima can be characterized as an early-stage deeptech vendor with meaningful but still limited market penetration. As of early 2025, CNRS cites a team of “about fifteen employees”; Seedtable lists company size as 11–51 employees, and Atoptima’s jobs pages show ongoing recruitment.²³⁴³²¹⁵ Funding consists of a single disclosed seed round plus public grants, with no later VC rounds or exits reported.¹³¹⁴³⁰¹⁵² Client references include some recognizable names (via Logtran and CMA CGM/CEVA), but the majority of case studies are anonymized, and the geographic footprint appears focused on France and nearby markets.¹⁶¹⁷¹⁸¹⁹²⁰²¹²²²³²⁴²⁵²⁷³³

Sales cycles are described by CNRS as long but promising, with each project heavily tailored to the client’s context and constraints—typical of B2B deeptech selling advanced OR solutions to operational management.² Overall, Atoptima is technically sophisticated but commercially modest: an appropriate fit for organizations with complex, high-value routing/packing/scheduling problems and the willingness to engage in collaborative projects with OR experts, rather than a plug-and-play solution for broad supply chain planning.

Conclusion

Atoptima is a technically strong, academically grounded optimization vendor that has translated decades of column-generation research into a suite of practical solvers for routing, packing, warehousing and scheduling. Its core competence lies in Julia-based decomposition frameworks (Coluna, BlockDecomposition) and the ability to formulate and solve large, block-structured mixed-integer programs relevant to logistics and industrial operations. Delivered via SaaS applications and an asynchronous orchestration layer (Galia), these engines produce prescriptive decisions—routes, loading plans, schedules—that can be embedded into existing TMS/WMS/ERP landscapes. Publicly verifiable client references such as Logtran (serving Carrefour-group banners), CMA CGM/CEVA and AppliColis demonstrate real-world deployment, albeit primarily in French and francophone markets and with many additional anonymized cases.

At the same time, Atoptima’s offer is narrow and deep: it does not encompass demand forecasting, probabilistic inventory optimization or end-to-end supply chain planning, and its “AI” claims are best understood as references to advanced optimization algorithms rather than to machine learning. Commercially, Atoptima remains an early-stage deeptech firm with limited headcount, a single disclosed funding round and a modest but growing client base. For organizations with challenging combinatorial problems—especially in routing and 3D loading—Atoptima’s technology likely offers state-of-the-art capabilities within its niche. For broader quantitative supply chain transformation, however, Atoptima would need to be combined with complementary tools or platforms (such as probabilistic forecasting engines) to cover forecasting, inventory policy design and multi-echelon risk management. The company’s evolution will depend on whether it continues to focus on being a specialist solver provider or broadens into a more comprehensive supply chain optimization stack.