Finance
NZ fintech (confidential)
LLM + structured extraction to automate financial report generation from unstructured documents.
faster reports
FTE freed
extraction accuracy
The problem
Analysts spent 3 days per report extracting data from PDFs, Excel sheets, and bank exports. 8 analysts × 3 days = 24 person-days per report cycle.
Our approach
Claude Sonnet for extraction with Pydantic output validation, pgvector for document lookup, and a Django API that generated structured JSON handed to a report template. DeepEval gate: 99% accuracy on golden set.
The outcome
Report generation dropped to 4 hours. 4 analysts reassigned to higher-value analysis work. Accuracy improved vs. manual process (human error was 2.1% baseline).
Ready to Grow Your Business Online?
Talk to our team today. We will understand your needs, suggest the right solution and give you a clear quote — all for free.