An investment research team at a global asset manager needed to process quarterly earnings reports across a portfolio of 200 public companies. Each report runs 80 to 150 pages. The team wanted to extract key financial metrics, management commentary on forward guidance, and any material risk disclosures — consistently formatted and ready for analyst review within hours of filing.

The Integration

The team built a processing pipeline using Crucible's API with a two-stage approach. In the first stage, Fast Mode was used to classify each section of the report and route it to the appropriate extraction template. In the second stage, Deep Mode processed the routed sections with structured schemas for each data type.

The two-mode approach kept costs manageable. Fast Mode handled the classification cheaply and at high throughput. Deep Mode was reserved for the sections where accuracy on ambiguous language mattered.

Output Format

Each processed report produced a standardized JSON object containing extracted financials, a summary of management guidance language, and a list of flagged risk factors with the relevant source passages attached. The JSON fed directly into the team's internal research platform.

What the Team Said

The analyst team reviewed Crucible's output against their own manual extractions on a sample of 40 reports. Agreement on quantitative fields was 97%. Agreement on qualitative summaries was rated as acceptable or better in 89% of cases. The two failure modes they identified were reports with non-standard layouts and cases where forward guidance was embedded in footnotes rather than the main body text.

Both failure modes were addressed in a subsequent prompt revision. Processing time for the full portfolio dropped from three days of analyst time to four hours of compute.