Unnamed Leading Litigation Finance Firm

A leading litigation finance firm built an AI system to identify and analyze U.S. federal court cases by downloading PDFs and extracting key facts via LLMs to accelerate funding decisions. The initial implementation suffered 90–95% error rates across data downloads, processing, and interpretation, rendering the platform unusable for downstream teams. Root causes included fragile browser automation, incorrect handling of delayed document availability, and dead-letter queue loops in Azure Service Bus.

SPR re-architected the platform end-to-end, redesigning court-site navigation to wait for full page loads before proceeding, introducing explicit 'PDF not yet available' status tracking rather than treating missing documents as errors, and adding case lifecycle logic to handle self-representation, duplicates, and closed cases. LLM error handling was hardened so missing PDF sections produced structured outcomes instead of crashes that fed the dead-letter queue. The underlying database schema was also redesigned from a denormalized structure to a properly normalized one.

Error rates dropped from approximately 90–95% to 10–15%, with the remaining errors attributable to source PDFs missing required sections. The platform was successfully deployed to production, enabling the client's content analysis and review teams to use it daily. The rebuild also gave the client's engineering team a stable foundation on which to continue feature development independently.

• Treating 'document not yet available' as a tracked lifecycle state rather than a system error is essential for pipelines that depend on external court document posting schedules.
• Dead-letter queue accumulation is often a symptom of unaddressed root causes; requeueing without fixing underlying issues compounds failures at scale.
• Re-architecting for real-world edge cases (delayed availability, case status variations, LLM gaps) requires a full rebuild rather than incremental patching when the foundation is structurally flawed.