Video: Starduct's AI assistants Lê Na & Gemma

Every day Starduct works with thousands of product codes, hundreds of projects, many languages — and engineering figures that must never be wrong. To work faster without losing accuracy, Starduct has built two AI assistants into its own digital platform.

Lê Na (Lena) — the communicator. Her full name, Đào Thị Lê Na, combines the names of four fruits typical of Vietnam. She talks with customers, engineers and distributors on the website in 5 languages (Vietnamese, English, Korean, Japanese, Chinese) and reports to the internal team.

Gemma — the processor. An open language model released by Google, running on Starduct's own server and built into the ERP that Starduct developed in-house.

Both share one source of data: Starduct's product database, catalogues and test reports.

Alongside people at every step of a project

• Learning about the product: Lê Na looks up every model, catalogue, test report and nearly 300 pages of technical content; quotation requests go to sales, difficult questions to R&D engineers.
• The offer: import a BOQ and technical documents — the system reads every line, selects the right model and code for each ordered size, and prepares the offer and quotation in Starduct's standard template with full technical data. The person in charge checks and approves. Try it: BOQ take-off.
• Design and production: parametric 3D drawings in FreeCAD, design checks against standards, models prepared for airflow simulation with OpenFOAM, material requirements and BOM, sheet nesting, production processes.
• Distributors and partners: orders, production progress and delivery status are updated on the Distributor Portal from ERP data; every question reaches the person in charge.
• In parallel: meetings, business trips, partner appointments and factory visits tracked; documents and reports drafted; at night, software and data reviewed — and more than 20,000 interface lines translated into English.

Four principles of trust

1. AI handles the language, software handles the numbers — the assistants never set prices or change figures.
2. AI drafts, people approve.
3. The right access, the right scope — payroll, cost and personnel data are out of reach; names are masked before anything reaches the model.
4. Test before release — Lê Na is scored against a standard question set and speaks about certifications only as far as the records allow.

Both assistants learn through their work, with human review before new knowledge is used — and they are still being optimised and equipped with new functions. Chat with Lê Na: the Ask Lena button at the bottom right of every page.