Innovate · Vision · Language · Predictive models
R&D
Does your problem call for mathematics? Let's look at it together.
Systems that read documents, recognise images, understand speech or anticipate an event. On your data, inside your IT system, without sending it anywhere else.
- Vision, document reading, speech and prediction
- Private AI, on site, integrated into your IT system
- Models delivered in an open format: you are not locked in
Why R&D
We say what we do, and we do what we say
Artificial intelligence is not magic: it is mathematics and a great deal of engineering. Everything comes down to three questions: is the problem well framed, does the data really exist, will we be able to tell that it works? We start there, and we say no when a subject cannot be treated seriously.
Your data goes in, a verified result comes out.
- Images
- Documents
- Voice
- Business data
A verified, explained result
What R&D does for you
Computer vision
Recognise and locate a defective part, a number plate, an area of a drawing.
Document reading
Invoices, delivery notes, forms: the image cleaned, the result checked.
Speech and language
Voice input on the device, with no external service.
Predictive models
A breakdown, a customer drifting away, an anomaly: every prediction explained.
How it works
From the problem to the model in production
- 1
Scientific scoping
What the system must do and how to check it, in black and white.
- 2
Data and exploration
Volume, quality, bias, gaps: no promises before we have looked.
- 3
Trials, from simple to complex
If the sophisticated model does no better, we deliver the simple one.
- 4
Deployment and integration
An open format, a user guide, the means to retrain.
On your own case
See R&D on your business.
A presentation with you: what R&D changes for you, what it takes, and how it starts. Our proposal comes afterwards, in line with what you have seen.
The full detail
For those who want to go deeper.
Case: reading administrative documents and invoices
- Clean the image: the document is straightened and cleaned.
- Find the useful areas: running text, tables, boxes to read.
- Read the text: a reading refined on your documents.
- Correct: errors fixed using the vocabulary of your industry.
- Check: anything still doubtful is flagged for review, not delivered as certain.
- Quality is measured on your documents, and we show you the errors.
Case: detecting objects in images
- A model already trained, then specialised on your images.
- It shows where each item is and what it is.
- A few hundred representative cases are enough, including the difficult ones.
- Validation is always done on images the model has never seen.
Case: predicting from business data
- Most of the work is on the data: what is missing, what is aberrant.
- Several families of models compared on your data.
- Every forecast is delivered with the factors that carried weight.
- We tell you when the available data cannot answer the question.
Tools
- Vision and image: OpenCV, Pillow, scikit-image, Tesseract 5, PyTorch, torchvision, ONNX Runtime.
- Audio and language: Vosk, Sherpa-onnx, WebRTC VAD, NLTK, spaCy, HuggingFace Transformers, Kaldi.
- Learning and data: scikit-learn, XGBoost, LightGBM, Optuna, SHAP, Pandas, NumPy, Matplotlib.
Who is this offer for?
- IT and technical departments: an outside view to scope or assess an AI project.
- Deeptech startups: moving from proof of concept to product.
- R&D project leaders: a technical partner at the frontiers of AI and mathematics.
Let's talk about your project.
A conversation to start, a single partner for what follows.