Open LLM hosted with you
- Open-weight models (Llama, Mistral, Qwen, DeepSeek)
- Deployment on your GPU server or on Tunisian infrastructure
- No data sent to an external service
- Performance on par with online services for business tasks
Service 2 — Artificial intelligence
Deploy AI without giving up your data
Lexa builds LLM agents hosted on your infrastructure or in Tunisia. They process your documents, answer your customers, automate your repetitive tasks — while your data stays with you.
Generative artificial intelligence has entered internal processes faster than any previous digital transformation. The reflex to 'just plug in ChatGPT' has a hidden cost: your prices, your contracts, your know-how go off to an American service that may use them to train its models. Local AI removes that risk without giving up the productivity gains.
Use cases we deploy
Internal conversational agents
An assistant that answers your employees' questions from your documentation, without exposing it outside.
Augmented search (RAG)
Your staff query your archives, contracts and technical files in natural language, in seconds.
Document automation
Automatic extraction, classification and summarisation of invoices, contracts, quotes, HR files, minutes.
Computer vision
Automated quality control in production, OCR of purchase orders, ID reading, barcode traceability.
Multilingual chatbots
A customer-service agent that answers in French, Arabic and English on your site, your WhatsApp or your inbox.
Predictive analytics
Sales forecasting, anomaly detection, route optimisation, predictive maintenance — on your own data.
The problems we address
The recurring pain points we see in Tunisian SMEs — and their concrete impact on the business.
Your queries go off to OpenAI, Google or Microsoft
Every prompt exposes your business data to a foreign third party. The terms of use often allow its use for training.
Regulatory exposure in sensitive sectors
Healthcare, finance, legal, defence: sending a patient file or a client contract to an online service doesn't hold up legally.
Dependence on a unilateral provider
Price, terms of use and availability can change overnight. You've wired your processes around a service you don't control.
Unpredictable cost per query
Once adopted, usage explodes. The monthly AI API bill becomes a recurring cost that's hard to cap.
Consumer chatbots don't read your data
A generic AI agent knows neither your catalogue, nor your order book, nor your history. Its answer is inevitably imprecise.
Our approach
RAG plugged into your content
- Indexing of your internal documents, archives, knowledge base
- Sourced answers that point to the original document
- Continuous update as you add content
- Permissions respected: each user only accesses their documents
Integration with your ERP
- Reading the catalogue, orders, stock, customer history
- Actions on the ERP (create a quote, change a status) from the agent
- Mixed human + AI workflows for sensitive tasks
Continuous quality measurement
- Tracking hallucination rates, accuracy, latency
- User feedback loop to improve the model
- Usage and business-ROI dashboards
Why Lexa for AI
Four concrete differences that matter when it's time to choose.
Absolute sovereignty over your data
All our agents run on a local LLM. No customer data, no internal document, no sensitive query leaves your perimeter. That's the condition for making AI acceptable in healthcare, finance, legal and industry.
Business use cases, not a tech demo
We always start from a high-value, measurable use case. The AI agent delivered serves a specific business process, not a technology showcase.
Modern stack on proven infrastructure
Llama, Mistral, Qwen, vLLM, LangChain, PyTorch vision: we work on the mature open-source ecosystem, not on closed proprietary frameworks.
Native articulation with Lexa ERP
When AI is plugged into your ERP, it accesses your company's real data and can act on it — a rare combination on the market.
Frequently asked questions — AI
The questions managers ask us most often on this topic.
A local LLM (local Large Language Model) is an artificial-intelligence model hosted on your own infrastructure, rather than on a foreign provider's servers. When your employees ask the agent a question or submit a confidential document, nothing leaves your company. This is essential for regulated sectors (healthcare, finance, legal) and for any company mindful of its digital sovereignty.
For most SME use cases, a dedicated GPU server is enough (typically an RTX 4090 or A6000-class card, with 24 to 48 GB of VRAM). For heavier loads or several dozen simultaneous users, you move to a multi-GPU workstation or shared hosting in Tunisia. The entry cost has dropped sharply and is no longer a blocker.
No, and that's not the goal. Well-deployed AI frees up time on repetitive, low-value tasks (information lookup, transcription, classification, standard first replies) so your employees can focus on high-value work (decisions, customer relationships, expertise). Experience shows AI increases productivity per employee rather than replacing jobs.
No. The model used is open-weight (Llama, Mistral, etc.) and runs locally. No customer data, no internal document, no user query is sent to the model's creator or a third-party service. You keep full control of your data and its use.
The right starting point isn't the choice of model or hardware, but identifying a high-value, measurable use case that justifies the investment. A few typical examples: a research assistant on internal documentation, automating the classification of incoming emails, a mail pre-sorting agent, quality control by vision in production. Lexa scopes this first use case for free during the initial audit.
Let's talk about your AI
Thirty minutes are enough to understand your context and identify the first opportunities. Free audit, no commitment.