Product · Enterprise Intelligence Assistant
Plain-English answers from your company's data.
The Enterprise Intelligence Assistant answers questions across your databases and documents together, explains its reasoning and shows the source of every answer. Its models are trained on your own documents, and it runs entirely inside your infrastructure.
No data leaves your environment.
The challenge
The answers are spread across systems.
The figures sit in databases and the context sits in documents and reports. Getting one answer often means writing SQL, searching files and reconciling the two by hand.
Split data
The numbers live in databases, and the explanations live in documents.
Unfamiliar language
General AI models have never seen your internal codes, entity names or report formats.
Answers you can't check
An answer without the query or passage behind it can't be verified.
Data that must stay put
Sensitive records can't be sent to a public AI service.
What it does
Three kinds of questions, one assistant.
A small decision model reads each question and sends it down the right path. When it isn't sure, it takes the combined path or asks you to clarify.
Figures from your databases
Questions become checked SQL that runs on your live tables. A failing query is repaired and run again.
Answers from your documents
A retriever trained on your own questions finds the passages that answer them, and the assistant summarises them.
Both, in one answer
Figures from the database and context from documents, combined into one reasoned explanation.
Sources on every answer
Each answer shows the SQL it ran or the passages it used, so anyone can check it.
Tables and charts
Results come back as a number, a table or a chart, whichever fits the question.
Always current
Databases and documents are ingested as they change, so answers reflect today's data.
How it works
Built on our 4-layer stack.
Every question runs the same fixed, auditable steps: route, retrieve, generate and verify. Models make the decisions inside each step.
Application interface
You ask in plain English and get text, a table or a chart, with its sources.
Knowledge fabric
Live ingestion into Postgres, with a graph of the schema, and into a vector store with source, date and access metadata.
Agentic intelligence
Route, retrieve, generate and verify, with failing SQL repaired and retried.
Sovereign AI models
One domain-adapted LLM plus three small models, each trained for one job.
The models
Trained on your own language.
Continued pretraining on your unlabelled documents teaches the model your codes, names and report conventions. Fine-tuning then teaches it five tasks. That is what lets a right-sized model do the job inside your infrastructure.
Domain-adapted LLM
Writes SQL, fixes failing SQL, builds charts, summarises documents and reasons over mixed evidence.
Decision model
Routes each question to the database, the documents or both, with a confidence score.
Retriever model
Trained on your own questions, so it finds the passages that answer them.
Entity model
Marks the entities a question names, so SQL is written against only the tables it needs.
Results
From 52% to 92% accuracy.
End-to-end, on held-out questions, the same pipeline answered 52% fully right with the base model and 92% after domain adaptation. Only the model's training changed.
| Measure | Base model | Domain-adapted model |
|---|---|---|
| End-to-end accuracy | 52% | 92% |
| Internal vocabulary | Not trained on internal documents | Learned through continued pretraining |
An LLM judge grades each final answer. A question counts only if the route, the data, the answer and its sources are all right.
Enterprise ready
Runs inside your boundary.
Designed for organisations whose data can't leave their own environment.
Your infrastructure
On your own GPUs, in your private cloud or air-gapped. No data leaves your environment.
Access control
Retrieval respects each user's permissions, with row-level security in Postgres and permission filters on documents.
Safe SQL
Queries run under a read-only database role and are validated before they run.
Audit trail
Every question is logged with its route and confidence, the SQL, the sources and the answer.
Guardrails
A guardrail model screens every question before it reaches the LLM and every answer before it reaches the user.
Improves with use
Query logs and expert corrections feed each training round, so accuracy grows with every deployment.
How to start
From your data to production.
The architecture, retrieval method and training recipe are reused. Only your schema, vocabulary and test questions are new.
Discovery
Map your data sources, collect real questions, and agree access and security.
Pilot
One domain: connect the sources, build the stores and measure a baseline.
Tune
Continued pretraining, fine-tuning, routing and retriever tuning, measured against the baseline.
Production
Deploy in your environment, with monitoring and the improvement loop.
Questions
Frequently asked.
What is the Enterprise Intelligence Assistant?
An AI assistant that answers plain-English questions across your databases and documents, explains its reasoning and shows the source of every answer. Its models are trained on your own documents, and it runs inside your infrastructure.
What data can it answer from?
Relational databases, and documents such as reports, policies and manuals. Both are ingested as they change, so answers reflect current data.
Where does it run?
On your own GPUs, in your private cloud or in an air-gapped network. No data is sent to an external AI service.
How accurate is it?
In our deployment, end-to-end accuracy on held-out questions rose from 52% with the base model to 92% after domain adaptation. We measure a baseline on your own questions before anything goes live.
Can we check an answer?
Yes. Every answer shows the SQL it ran or the passages it used, and every question is logged with its route, SQL, sources and answer.
What do you need from us to start?
An inventory of your data sources, real questions with expected answers, a domain expert for labelling, and your target infrastructure and security requirements.
Contact
Tell us what you need.
We partner with a limited number of enterprises at a time, so every engagement gets senior attention. Tell us what you want to build, and our leadership will reply.