Sylar – Expert AI | Meeting Airbus 01/15/2026

AI landscape - state of the art

the expert behaviour self learning :

Model is growing during its exchanges with end users. Thanks to methodology applied and efficiency of dataset created, first trials have shown :
- 200 user exchanges  50 quality patterns  2 weeks of usage (averages) shown a change in agent behaviour
- 1000 user exchanges  200 quality patterns  more or less 2 month of running shown expert corrections drop dramatically, accuracy of answer is close to 80%
Day 1 model propose generic solution and knowledge on main pinciples
Day 60 model is able to provide an accurate methodology as an expert!

main points

2 main players categories on the market :
big players and Small / medium actors
Great assistant, but far from an efficient Expert
How to have a solution to train and use expert at same time?
How to have a self eveolving system?
Different expertise but same way of thinking process (measure, identify, incrental implement)
Expert = knowledge + pattern of thinking
new era off Agentic agent open in 2026.
We were able to implement a cascade of models
(from 4b up to 81B parameters) with dedicated functions for each ones.
a full model of 70b parameters is dedicated
only to identify and formalized the pattern of thinking of the request sent,
validate and integrate user feedback,
correct it and certify this way of thinking,
under control and validation of human expert!
dedicated reasoning model is constrain to apply it by our agent orchestrator,
on verification loop is applied and methodology is checked for each answer,
once a complexity agent detect the needs
periodically (weekly or daily at the beginning, monthly later,
an integration of VALID and CERTIFIED partetern in the model weitgh are applied,
to integrate natively the new pattern inside the reasoning model.
Expert = knowledge + pattern of thinking

Our solution -- Inside the brain

Instead of a big model with large context
2 works has been done to optimize solution foot print :

Mixture of models
Agentic modulable agent architecture of the solution is built to allow each llm focus on simple task.
Communication language between them is encode to ensure low data transfer intra memory

1 llm focused on natural language, 1 llm focused on request complexity, models focused on memory management and promotion

This allow us to reduce our core agent footprint and allow a run on previous GPU generation (Ampere archiutecture)
(test shown latency of 40s for 1000 Tokens on 2 A100 config with 5 concurrent requests).
Second optimization has been on memory and knowledge management.
Using a memory store with dedicated “archivist” supplied by embedding and ranking models,
plus the development of vectorized database allowed us to optimize and control the context passed to the model.
An evaluation of memory has shown 94% success to LoComa tests (not certified).
Expert = knowledge + pattern of thinking

Medium llm with lower precision + context controlled == scalable on smaller server.
designed for small units deployment.

What's next?