Virtua Intelligence

Historical — what actually happened. Reported figures across trailing periods.

Forecast — what is expected. Forward-looking estimates and projections.

Benchmarking — how it compares. Peer and comparable-set data for relative analysis.

Documents — why, and what was said. Unstructured content extracted to give research context the tables cannot.

Financial data is only as accessible as the query language standing in front of it. Reckonsys built a natural-language layer over Virtua's data estate: an analyst asks in everyday language; the system works out which datasets the question touches, retrieves across them, pulls in supporting context from documents, and returns an answer. No SQL, no schema knowledge, no waiting on someone who has both.

Everyday-language questions are translated into retrieval across the financial datasets, without the analyst writing or seeing a query. A single question can touch several datasets, with results combined rather than returned separately. Dataset categorisation gives the system a basis for deciding which sources a question actually needs.

The hard part of natural-language querying in finance is rarely the language. It is that an analyst's question implies a set of datasets and a way of relating them, and none of that is stated in the question. Categorising the estate first is what turns an ambiguous request into a resolvable one.

A question can be answered with figures and with the commentary that explains them, in one response. Research context that exists only in prose becomes reachable through the same interface as the numbers.

This is the difference between a query tool and a research tool. Answering both from one question is what removes the manual step of going to find the filing.

Built for analysts, not engineers. There is no requirement for technical knowledge of databases or query languages. Exploration becomes iterative — a follow-up question rather than a new query to write — and the people with the domain questions can reach the data directly rather than through a request queue.

Claude sits at the point where an analyst's question meets a set of databases that know nothing about how analysts talk. It interprets what is being asked, determines which of the categorised datasets the question actually requires, and assembles the retrieved figures and document context into a single answer.

Margin, growth, peer group and comparable all carry precise meanings that shift with context, and a question phrased casually still implies a specific calculation over specific periods against a specific comparison set. Getting from that to the right retrieval is a reasoning problem, and it is the one Claude is doing.

The assembly step matters just as much. Combining a forecast figure, a peer median and a line of management commentary into something an analyst can act on requires holding several sources in view at once and being careful about what each one actually supports. That is where a capable model earns its place over a query translator.

Stack: Claude API · natural-language to retrieval · multi-dataset querying · document extraction