Build agents that actually run the business — on our own hardware, not someone else's cloud.
We run our own data centre and our own Odoo 19 ERP. Quotations, invoices, interventions,
tickets: it all lives here, in Tunisia, on machines we can point at. That makes InTech an unusually
good laboratory for local agentic AI — models that read and act on real operational data,
without a single row leaving the building.
The project
You will help build an agent layer over our ERP: a set of tools an open-weights model can call
through MCP (Model Context Protocol) to answer questions and do real work — draft a quotation from
an email, flag invoices about to age past due, prepare an intervention sheet, summarise a customer's
history before a call. We already have an MCP server talking to Odoo; your job is to make it
genuinely useful, and safe.
What you will actually do
- Run open-weights models locally and find out, honestly, which ones are good enough for the task.
- Design tools an agent can call, and the guardrails that stop it doing something expensive by accident.
- Build retrieval over our own documents: quotations, datasheets, past interventions.
- Measure it. "It felt better" is not a result — we want evaluations we can repeat.
- Ship something a colleague in sales or support uses on a Tuesday without being asked to.
What we are looking for
- You write Python without needing to be supervised, and you read documentation before asking.
- You have played with LLMs beyond a chat window — APIs, local inference, prompts you iterated on.
- You are careful with other people's data, because this is production and it is real.
- You can say "this does not work" about your own idea. We value that more than enthusiasm.
- Final-year engineering / master's, or equivalent evidence you can build things.
Not required: a paper record in machine learning, an existing Odoo background, or perfect English.
French or Arabic is fine — curiosity is the part we cannot teach.
Practicalities
- 4 to 6 months (PFE welcome), based at our La Marsa office, some remote days.
- Mentored by the team that runs the infrastructure and the ERP — you will see production, not a sandbox.
- Paid internship, with a real chance of joining us afterwards if it goes well for both sides.
To apply: send anything you have actually built — a repo, a script, a broken prototype with a
good story. That tells us more than a CV does.
