All projects
LiveAI & Intelligent SystemsDetails pending
AI Jack
A self-hosted AI assistant that answers questions about a company's own ERP data — embedded as a chat widget inside existing business applications, running on local models so no business data leaves the server.
- Type
- Type to confirm
- Category
- AI & Intelligent Systems
Technology
- Python
- FastAPI
- Ollama (local LLM)
- ChromaDB
- Sentence-Transformers
- MySQL
- Redis
- Embeddable widget
Challenges
The problem
Business users keep asking the same data questions of an ERP — what is pending, what changed, who owes what — and each answer needs someone who knows the schema. Sending that data to a third-party AI service was not an option, and the existing applications were built in different stacks.
Architecture
Shape of the system
- Widget → API. A small embeddable widget in the host application talks to a FastAPI service.
- Local model. Inference runs on a local Ollama instance; there is no dependency on an external AI API for the core path.
- Grounding. Questions are answered against the live relational database plus a ChromaDB vector store (sentence-transformer embeddings) built from schema knowledge and documents.
- Safety. The service is a separate process behind the reverse proxy; the host application only passes a token and renders the response.
Solutions
What it does
- Natural-language questions over ERP data, with schema-aware query planning.
- Report/Excel/PDF export of answers.
- The same assistant is reused as an embedded helper in other internal apps (recruitment, distribution) via a thin proxy.
- Built to run without paid APIs, with the model swappable.