What it does
The system turns raw financial filings into structured working material for analysts and finance teams. Judgement stays with the analyst or finance team. The machine handles the mechanical work of assembling a foundational financial analysis, while surfacing tensions that warrant further diligence, saving hours to days of work that would otherwise go into pulling the material together from raw filings.
The system is designed to minimise LLM accounting errors, misguided associations, and outright hallucinations. The system is vendor-agnostic, allowing individual tasks to be optimised for the model that delivers the best combination of accuracy, reliability and cost.
The underlying patterns and hypotheses that guide the financial analysis can be tailored to reflect a firm's own analytical approach.
Who it's for
For an analyst, it works as the foundation for pre-investment research and context-based portfolio monitoring. Identified tensions are translated into evidence-based diligence questions, giving the analyst a starting point for further investigation or management engagement. New filings and disclosures can be assessed against the existing evidence base, allowing the analyst to see what has changed, whether previous signals have strengthened or weakened, and where further investigation may be warranted ahead of the next investment or portfolio decision.
For a CFO or internal finance team, new information can be incorporated as it becomes available, offering a fast, systematic way to see what is changing, what needs explanation, and what evidence is needed before an issue becomes a question from the board, lenders or investors.
How it's built
The LLM is treated as a bounded inference component within a wider, modular system. Tasks that can be performed reliably in code are kept outside the LLM. Model calls have specific jobs and constrained inputs, with deterministic validation, machine learning-based optimisation and LLM-based evaluation around their outputs.
The result is evidence-based analysis that is repeatable, testable and updateable as new information comes in, with a reduced probability of error.
Read more about the architectural approach and design principles →