[PREDICTIVE MODELS]

Your bespoke predictive model, trained on your data

We build and train predictive models on your historical data, then integrate them into your operational tools. You anticipate demand, an equipment failure or customer behaviour, and make better decisions earlier.
50+ predictive models in production and paying their way for our clients
Up to 98% forecast accuracy with our models
140+ senior tech profiles from France's top schools: Polytechnique, Centrale and ENS
Talk to an expertSee 5 predictive models
Diagram of a neural network: the predictive model trained on your data
800+ organisations trust us with their tech and AI projects
[OUR SPECIALITIES]

Example applications

We train models on your data, integrate them into your operational tools and maintain them in production, with a high level of accuracy
Demand and sales forecasting

Demand and sales forecasting

Time-series models project your future activity from your history and the factors that drive it: weather, the school calendar or key trading periods
Financial and cash-flow forecasting

Financial and cash-flow forecasting

Forecasting models project your revenue, cash receipts and cash position over your financial year, with a confidence interval for every projection
Predictive maintenance and failure forecasting

Predictive maintenance and failure forecasting

Maintenance models learn the signals that come before a failure on your equipment and trigger an intervention before production stops
Stock and capacity optimisation

Stock and capacity optimisation

Forecasting models size your purchasing, staffing and capacity to the demand you can actually expect, day by day and site by site
Missing-data estimation and reliable KPIs

Missing-data estimation and reliable KPIs

Estimation models reconstruct the readings your sensors or meter reads failed to report, so your contractual and regulatory KPIs hold up even when only part of your estate is instrumented
Lead time and duration forecasting

Lead time and duration forecasting

Forecasting models estimate a delivery lead time or a production duration from your history and the day's conditions, and raise an alert as soon as a deadline becomes unrealistic
Fraud and anomaly detection

Fraud and anomaly detection

Classification models learn what is out of the ordinary in your transactions, flows or readings, and flag suspicious cases in real time
Scoring, churn and customer segmentation

Scoring, churn and customer segmentation

Scoring models give each customer a probability of leaving, defaulting or buying, so your teams know which accounts to prioritise
Recommendation and dynamic pricing

Recommendation and dynamic pricing

Recommendation models work out the most relevant product and price for each customer at every point in your trading season
Marketing performance and attribution

Marketing performance and attribution

Regression models link your sales to the marketing activity that generated them, and estimate the expected return on a budget before you commit it
[Our work]

Five predictive models in service with our clients

SCHOOL CATERINGTurboself

Forecasting canteen attendance to cut food waste

Turboself equips schools with payment and catering management solutions. Its clients ordered food without knowing how many pupils would turn up for lunch, and threw away the surplus. Galadrim developed the attendance forecasting algorithm built into its software. The model is trained on historical attendance, data from Météo France (the French national weather service), teacher attendance and menu analysis by a language model. Each canteen adjusts its orders to the day's expected attendance, and the model retrains as new data comes in.
5 500client schools
2Mdaily users
÷5food waste costs
Turboself Logo
DISTRIBUTION NETWORKSSuez

Estimating water meter consumption for a reliable network efficiency figure every day

Every day, Suez reports to its local authority clients the efficiency of their water network, an indicator that depends on knowing the consumption of every meter. Part of the meter estate has no remote reading or sends irregular data, which undermined the indicator's accuracy. Galadrim developed the estimation engine that fills these gaps. An XGBoost model predicts consumption for meters with a history, taking weather and local correlations into account. A statistical method builds standard profiles for sparsely equipped estates. Monte Carlo simulations then consolidate everything into an efficiency figure with its confidence interval.
600 000residents covered
59 000meters with reliable data
95 %accuracy
Suez Logo
FinanceBNP Paribas

Strengthening the algorithm that builds investment portfolios for two banks

BNP Paribas and Hello bank offer their customers a portfolio of shares and money market funds, built from a risk appetite questionnaire. Galadrim maintains and develops the algorithm behind these recommendations: checking input data, calculating input variables, handling error cases and optimising response times to meet the required thresholds. An adviser gets a proposal consistent with the profile their client has declared, and monitoring tools flag any drift before it reaches customers.
+200 000portfolios built / year
<2 sto build a portfolio
BNP Paribas Logo
INTERNATIONAL FREIGHTLogfret

Assigning a product's customs code from 5,000 classifications

Logfret is an international freight specialist. Its clients classify every exported product under the HS customs nomenclature, choosing from 5,000 codes, and that code determines the tariff applied. Galadrim developed the classification engine that assigns the code from the product description alone, drawing on every declaration already processed. Teams get an instant suggestion to validate, on every declaration line.
5 000+customs categories
65countries covered
300employees
Logfret Logo
PUBLIC PROCUREMENTOrdiges

Pre-filling public procurement files with prediction

Ordiges publishes Lia, the software government bodies, local authorities and hospitals use to prepare their public tenders. That means dozens of regulatory fields to fill in and a European classification code to choose, even though very similar tenders already exist. Galadrim built the software's two prediction algorithms. The first finds similar tenders and suggests field content to reuse. The second suggests the classification code that matches the subject of the contract. Each buyer's data stays ring-fenced, and reused files are anonymised before they are suggested.
+80 000indexed references
40+ areas of expertisein public procurement
Ordiges Logo
[Our approach]

How we put a predictive model into production

  • Scoping
    Step 01

    The decision to support, the target indicator and the accuracy threshold

    We start from the decision the forecast has to serve, and agree with you on the indicator to predict and the acceptable level of error. That threshold becomes the go-live criterion, set before any development begins.
  • DATA
    Step 02

    Auditing and preparing your historical data

    We audit your existing data and measure its depth and its gaps. Then we prepare it and enrich it with external sources that improve the forecast: weather, calendar or market data.
  • MODELLING
    Step 03

    Choosing the model and training it on your data

    We choose the right approach for your problem (time series, classification or regression), then compare models on a test set that you approve. These models are deterministic: the same data gives the same forecast, and the calculation can be reproduced at any time.
  • INTEGRATION
    Step 04

    Forecasts delivered where decisions are made

    Our engineers surface the forecasts in your operational tools rather than in a separate dashboard, and integrate them securely and robustly into your existing IT systems.
  • DEPLOYMENT
    Step 05

    Go-live and variance tracking

    We make the calculation pipeline production-grade, automate its execution and set up monitoring that compares every forecast with actuals. Any variance shows up the moment it appears.
  • longevity
    Step 06

    Retraining and continuous improvement

    We retrain the model on new data and adjust it when your business changes, so it stays accurate year after year.
01/06
[Why Galadrim?]

Why do companies trust us with their predictive models?

Proven expertise

800+ clients, 170+ experts and 9 years' experience building AI products, with predictive models in service across industry, public services and banking.

Models trained on your own history

We audit your data, fill its gaps and enrich it with the external sources that really explain your variations. Accuracy is measured on your own data before anything goes live.

Auditable, deterministic forecasts

Given the same data, a predictive model always returns the same forecast, and every forecast can be traced back to the variables behind it. You can justify your decisions to a board, an auditor or a regulator.

Forecasts built in where decisions are made

A model delivered on its own is just a demo. We build the tool that puts the forecast in front of your teams when they order, plan or make trade-offs.

Accuracy monitored after go-live

We compare every forecast with actuals, retrain the model on new data and adjust it as your business evolves, so it stays reliable over time.

Let's bring your project to life together

Contact us
[Our team]

A team that advises you before it builds

We recruit and train passionate engineers and experts from the best schools, who combine AI expertise with software engineering skills. That hybrid profile lets us design your bespoke AI systems and take them all the way into production.
Benjamin Drighès
Benjamin Drighès CTO Data & AI
Advises our clients on product strategy, long-term AI transformation and key choices of technology and approach, and sets the quality standards for our AI teams.
Quentin Massonnat
Quentin Massonnat AI Team Lead
Designs the architecture of predictive computation pipelines, including data preparation, training and automated execution, and oversees development through to go-live.
Marc-César Garcia-Grenet
Marc-César Garcia-Grenet AI Team Lead
He ships the features that bring a forecast into your teams' tools, from the model to the application around it.
Félix Monnier
Félix Monnier AI Team Lead
Trains and compares forecasting models suited to your data, then takes the one that meets the agreed accuracy threshold into production.
Lucien Maillard
Lucien Maillard Principal AI Strategy Consultant
Scopes predictive projects with client leadership teams, identifies the decision to support and assesses feasibility against the data available.
Pierre-Antoine Dornic
Pierre-Antoine Dornic Principal AI Strategy Consultant
Maps AI use cases, prioritises them by ROI and feasibility, shapes leadership teams' roadmaps and steers deployment through to adoption by staff.
Eva-Garance Tison
Eva-Garance Tison Lead AI Product Manager
Turns business needs into AI products people actually use: scopes projects with clients, prioritises the roadmap, steers team delivery and makes sure what we deliver has real impact.
Naïs Schietecatte
Naïs Schietecatte Senior AI Product Manager
Runs the product cycle for AI solutions, from functional specifications to the control interfaces used day to day, working closely with engineers and business users.
[FAQ]

Questions our clients often ask

Your history on whatever you want to predict, in whatever format it already exists, including exports from your ERP or business software. We audit it to establish what it can predict. Then we handle its preparation and enrich it with the external sources that explain your variations.
We aim for the best level achievable on your data and check it on a test set before anything goes live. Our models in production exceed 95% accuracy on their use cases.
The decision stays with you. Every forecast comes with its confidence interval, and your teams can adjust it whenever they have information the model doesn't.
Yes. A predictive model is deterministic and auditable: the same input always produces the same output, and we keep a record of the variables that weighed on each forecast. So you can back up a decision in front of a committee, an auditor or a regulator.
We continuously compare every forecast with actuals and show the variance in your tool. As soon as accuracy slips, we retrain the model on recent data.
You do. Code, trained models, prepared datasets and documentation are yours at every delivery. The transfer of rights is written into the contract from day one.
We size every project to your scope and priorities. Tell us about your context and we'll come back to you within 24 hours with an initial assessment.
[Contact us]

Let's bring your project to life together

We work with every kind of client, across every industry. Whether you are an entrepreneur or lead a large organisation, we put together a team that fits your need.

More than 800 companies have trusted us to build their web, mobile and AI products

Your request

We'll get back to you within the hour.