AI Integration (Python + LLM APIs)

The challenge

Managers wanted quick answers about sales, stock and finance without waiting for custom reports. An earlier in-ERP assistant had serious problems:

I audited the earlier assistant and documented each defect.
I designed a new assistant around permissions and confirmation, not prompts. The model never sees data the user couldn't see in the ERP, and it can't act on its own.

Provider abstraction in Python (requests): OpenAI, Anthropic, Gemini, Ollama (local) and any OpenAI-compatible endpoint.
Pipeline: intent router, then permission service, then tool executor, then prompt/context builder.
38 ERP tools across sales, inventory, accounting, CRM and projects. They are read-only, and any action requires explicit user confirmation.
Guardrails: payroll data is never sent to a provider; API keys are redacted from error messages; model "thinking" blocks are stripped.
UI: an OWL chat systray and dashboard.

Calling the Gemini API from an Odoo module: fixed a client-lifecycle bug and moved the API key to an environment variable. The real API returned recommendations and 3072-dimension embeddings.
Running local LLMs with Ollama (general and code models).

59/59 automated tests, including a browser tour, run with a mocked provider on a database copy. Figures were cross-checked against the ledger on that copy.
Not in production. It is presented here as a tested prototype.