Enterprise / Sovereign AI
Enterprise Intelligence Assistant for Live Data and Documents
Client A US enterprise
For a US enterprise, Jungroo AI built an assistant that answers plain-English questions across a real-time stream of structured data in databases and unstructured data in documents and reports. It runs inside the client's environment and shows the query or passage behind every answer. On held-out questions, end-to-end accuracy rose from 52% with the base model to 92% after domain adaptation.
The challenge
Teams needed answers from two kinds of live data: tables that change through the day, and documents and reports. A general open-weight model answered only 52% of their questions fully right. It didn't know the company's internal codes and terms, and the data couldn't leave the company's environment.
The solution
Jungroo AI built one assistant on its four-layer sovereign stack, inside the client's environment:
- Live Ingestion: Databases and documents are ingested as they change: tables into Postgres with a graph of the schema, and documents into a vector store with source, date and access metadata.
- Question Routing: A small decision model sends each question to the database, the documents or both, and asks the user to clarify when it isn't sure.
- Schema-Aware SQL: An entity model marks what the question names, and GraphRAG walks the schema graph, so the model writes SQL against only the tables it needs. A failing query is repaired and run again.
- Domain-Tuned Retrieval: A retriever trained on the company's own questions finds the passages that answer them.
- Domain-Adapted Model: Continued pretraining on the company's own documents, then fine-tuning on five tasks: writing SQL, fixing SQL, building charts, summarising and reasoning.
- Cited Answers: Every answer shows the SQL it ran or the passages it used, and every question is logged with its route, sources and answer.
The impact
Impact
| Measure | Before | After |
|---|---|---|
| End-to-end accuracy | 52% with the base model | 92% after domain adaptation |
| Internal vocabulary | Not trained on internal documents | Learned through continued pretraining |
Only the model's training changed between the two results: the same pipeline, retrieval and held-out questions were used for both. An LLM judge graded each final answer, and a question counted only if the route, the data, the answer and its sources were all right.