Califrais / Our DNA
Science, tested in the field.
Califrais is a French deeptech founded in 2014. For more than ten years we have been developing algorithms for the food supply chain with France's leading academic laboratories, and testing them every day on real flows.
Manifesto
We believe the food supply chain is a lever for transformation. For too long it has been invisible, fragmented, wasteful and carbon-intensive.
It is time to take artificial intelligence from promise to proof, through rigorous research connected to the real world.
We build the AI tools that model the logistical complexity of fresh food, orchestrate flows and reduce the carbon footprint of every kilometre, every order, every delivery.
Because in the 21st century, waste is a collective failure.
- BoldnessWe embrace the ambition of transforming a vital, complex and historically locked sector.
- RigourAt Califrais, AI is not a buzzword but a lever. Every line of code targets a concrete optimisation; every project puts our scientific credibility on the line.
- ImpactOur responsibility is environmental, social and societal: our solutions meet the economic and ecological needs of every actor willing to improve.
Our story
Twelve years in the making.
- 2014
Califrais is born
Califrais is born with one idea: simplify and pool fresh-food procurement.
- 2018
One order for the whole market
A procurement platform for food-service professionals, and the first pooled delivery routes.
- 2020
Fresh food delivered to households
During lockdown, Califrais sets up a fresh-food delivery platform for households within days.
- 2021
Official operator of Rungis
Califrais becomes the official digital and logistics operator of Rungis and launches its innovative solution, RungisMarket.com.
- 2022
A joint laboratory with public research
The LOPF joint laboratory is created with CNRS and Sorbonne Université.
- 2023
Logistique 4.0 award
Winner of ADEME's call for projects under France 2030, to industrialise our technologies across the French wholesale market network. First papers at NeurIPS and ICML.
- 2024
A year of publications
Papers at NeurIPS, ICML, Biometrics and the Journal of Nonlinear Science.
- 2025
€10M Series A
A funding round announced at VivaTech to accelerate the decarbonisation of the food supply chain with AI.
- 2026
Research in the spotlight
The French Minister for Higher Education and Research visits Califrais and its academic partners. Paper at ICML 2026.
The context
The food supply chain faces three structural uncertainties.
Three trends make replenishment logic based on repeating last year obsolete. We must anticipate now: decarbonising the supply chain is becoming both an ecological and an economic imperative.
×5
Weather disasters in 50 years
Droughts, heatwaves, extreme events: a direct impact on harvests and logistics.
↗↘
Demand volatility
Consumer behaviour no longer repeats from one year to the next.
CO2
Tightening regulation
Existing and upcoming rules on industry emissions.
Two essential responses
What the food supply chain must be able to do, starting today.
Adapt in real time
Use every available variable, far beyond last year's history.
Optimise the whole chain
Make every kilometre driven and every kilo stored count: model suppliers, warehouses and points of sale together.
Which kind of AI
Generative AI is not enough. You need specialised AI.
Large language models are not suited to food supply chain problems: rather than a black box with exploding budgets and collapsing performance, the sector needs control and explainability.
General-purpose generative AI
- Not built to optimise systems with millions of variables and constraints
- No guarantee of respecting physical and commercial constraints
- Plausible answers, without bounds or guarantees, that do not necessarily respect real-world physical constraints
- Hard to audit, very costly at scale
Califrais specialised AI
- Machine-learning forecasting: high dimension, adaptive algorithms, continuous learning
- Operations-research optimisation: multi-objective, multi-level, robust to uncertainty
- Hard constraints respected: cold chain, use-by dates, capacities
- Interpretable and sovereign: every decision is auditable, interpretable and built on 100% proprietary technology
Our platform's agents use language models where they excel, for human-machine and agent-to-agent interfaces, and leave decisions to proven optimisation algorithms.
Our research model
A joint laboratory, an academic ecosystem.
Our model relies on structured cooperation and a unique close-knit set-up with academia: 12 laboratories involved, PhDs carried out jointly between the company and universities, post-docs, research engineers. Our researchers publish under their Califrais affiliation at the leading AI conferences.
The LOPF joint laboratory, for Large-scale Optimization of Product Flows, brings Califrais together with the Laboratory of Probability, Statistics and Modelling (CNRS, Sorbonne Université, Université Paris Cité), with support from the French National Research Agency.
Our work sits at the interface of machine learning, operations research, reinforcement learning and combinatorial and robust optimisation.
- 200+Researchers and professors involved
- 15PhDs in AI on our teams
- 13Publications since 2023, including ICML and NeurIPS
Our research
Four major research programmes.
Each programme starts from a real problem met in the field and aims for a publishable, then deployable, result.
Routing and pooling
Deciding every day how many vehicles to commit and in which order to deliver: fewer trucks, less CO₂, higher load rates, no late deliveries.
- Ongoing2023–2026PhD or post-doc
Stochastic optimisation and generative models for vehicle routing
Theoretical guarantees for deep generative models, extended to capacitated routing with time windows on real industrial data.
With Sorbonne Université
- Ongoing2025–2028PhD or post-doc
Hybrid learning and combinatorial optimisation strategies
Combining combinatorial solvers and learning for dynamic routing, with statistical guarantees on the full pipeline.
With Sorbonne Université
- New2026–2029PhD or post-doc
Routing with collaborating actors
One fleet per actor plus a shared fleet: discrete optimisation and game theory to pool deliveries across actors with sometimes competing interests.
With École des Ponts ParisTech, Institut Polytechnique de Paris
Demand forecasting
Forecasting demand for thousands of fresh products: high dimension, sparse data, seasonality, regime changes. It is the building block for replenishment and pricing.
- Ongoing2025–2027PhD or post-doc
Hierarchical forecasting with operator-valued kernels
Kernel methods and deep learning for multi-product forecasts that respect the catalogue's hierarchy, with statistical guarantees.
With Sorbonne Université, Université Paris-Saclay
- Ongoing2024–2027PhD or post-doc
Adaptive multi-horizon probabilistic forecasting
Kalman filters with stochastic covariates, expert aggregation and application to hierarchical data.
With Sorbonne Université
Inventory and replenishment
Deciding every day which products to order to meet demand while minimising cost and waste, despite uncertain sales, perishability and supplier constraints.
- Completed2021–2024PhD or post-doc
Online learning for supply chain problems
Regret bounds for stateful inventory problems, including perishability, order lead times and warehouse capacity.
With Université Paris Cité
- Ongoing2025–2028PhD or post-doc
Bandits and reinforcement learning for wholesale market logistics
Online methods with discrete decisions for non-i.i.d. inventory dynamics.
With Université Paris Cité
- Ongoing2025–2027PhD or post-doc
Lot-sizing with minimum order quantities and perishability
Discrete and stochastic optimisation algorithms for fresh-produce market replenishment.
With Le Cnam, Sorbonne Université
- New2026–2029PhD or post-doc
Integrated inventory and routing optimisation for wholesalers
A multi-level model combining replenishment modes, supplier pickups and customer deliveries.
With Le Cnam, LIPN
Market and customer dynamics
Understanding wholesale markets as dynamical systems: seller attractiveness, buyer behaviour, customer loyalty.
- Completed2021–2024PhD or post-doc
Survival analysis and churn prediction
Using many longitudinal variables to predict customer churn in real time.
With Université Paris Cité, Inria
- Completed2022–2025PhD or post-doc
Modelling the dynamics of fresh-produce markets
A dynamical system of wholesaler populations, analysed theoretically and validated on real market data.
With Université Paris Cité
Publications
Published at leading conferences and journals.
IEEE Trans. Signal ProcessingReadRobust Estimation for Linear State-Space Models with Stochastic Covariates
ICML 2026ReadLatent Guided Sampling for Combinatorial Optimization
NeurIPS 2025ReadWasserstein Convergence of Critically Damped Langevin Diffusions
AISTATS 2025ReadTheoretical Convergence Guarantees for Variational Autoencoders
- NeurIPS 2024Read
Non-asymptotic Analysis of Biased Adaptive Stochastic Approximation
- ICML 2024Read
Dynamic Survival Analysis with Controlled Latent States
Biometrics 80(4)ReadAn efficient joint model for high dimensional longitudinal and survival data via generic association features
J. Nonlinear Science 34(5)ReadDynamics of Buyer Populations in Fresh Product Markets
- SpringerRead
The insertion method to invert the signature of a path
PreprintReadOnline Policy Selection for Inventory Problems
- NeurIPS 2023Read
Online Inventory Problems: Beyond the i.i.d. Setting with Online Convex Optimization
- ICML 2023Read
Learning the Dynamics of Sparsely Observed Interacting Systems
- PreprintRead
Multivariate Online Linear Regression for Hierarchical Forecasting
Ecosystem
They advance research with us.
They support us
Join the scientific adventure.
Researchers, engineers, PhD students: we hire demanding profiles who want to see their models run in production.