The platform · lalla.ai
lalla.ai is the AI platform beneath everything Innopas ships, models, retrieval, agents, evaluation, guardrails, and Lalla Chat as the surface people actually touch. A new product brings a domain graph and an interface. It does not bring its own AI stack, and it never brings its own research.
The stack
This is the whole economic argument. Research proves a method once, the platform carries it once, and each new product pays only for its own domain and its own screens.
The test of a platform is the second product. Anything can be called a platform while one product runs on it. TopSyllabus and EUNIQ share no customers, no regulator and no vocabulary. And they run on the same layer two.
Inside the platform
Six capabilities, versioned and shared. A product team consumes them; it does not reimplement them, and it cannot quietly fork them.
Retrieval that walks a typed domain graph rather than searching a pile of documents. The answer is assembled from the structure, so it can name the node it came from.
Agents that query the graph, run a model, call a service and compose a result, and with each step recorded, so what happened can be reconstructed later.
Every output carries its supporting evidence and a confidence value, produced at the same time as the answer rather than reconstructed afterwards.
Test sets, baselines and regression runs for every capability. A change ships when it beats the previous version on the record, not when it demos well.
Access control down to the graph node, redaction, refusal behaviour and audit logging. A student and an operator see only their own scope, enforced below the product.
Model-neutral by design, and deployable in your cloud, on-premise or in a sovereign environment — because some of our customers cannot send data anywhere else.
Lalla Chat
Lalla Chat is the same component in both products. What changes is the graph it is grounded in and the scope it is allowed to see. This is the reuse argument, shown rather than claimed.
Lalla Chat walks the utility graph from feeder to transformer to meter, finds that metered consumption fell while load held steady at DT-024, and returns the pattern with its evidence window and a confidence value. The operator can open the node it named.
Lalla Chat walks the concept graph from the question back through its prerequisites, finds four answers that describe the effect but never the driver, and points one concept upstream to water potential: with the answers it drew that from.
Same component, same behaviour. Ground the answer in a graph, name the node, show the evidence, state the confidence. Only the vocabulary belongs to the industry.
How a question moves through it
The order carries an argument. Scope is applied before anything is read rather than as a filter on the way out, the difference between a permissions model and a disclaimer.
From Lalla Chat, or from a product API on a schedule. Same path either way.
Guardrails decide what this user may see before anything is read.
Retrieval follows real relationships, feeder to transformer, concept to prerequisite.
Agents run the method DTRIHub proved, using whichever model fits the task.
The nodes used and a confidence value ship with the answer, not after it.
Both routes are listed under Products, with what you get and who buys. Most customers meet the platform inside EUNIQ or TopSyllabus and never think about it again. The platform has its own site at lalla.ai, and the people who build it are named.
Start here
Thirty minutes with the engineers who would do the work. No deck, no discovery invoice, a straight read on feasibility, sequence and what a first build would take, including when the answer is that you should not build it.
