Production work that proves the thesis

Challenge

Broker reports hold critical data, but formats vary and standard extraction misses the financial semantics.

Approach

Meta-prompting + PydanticAI, designed by someone who has read hundreds of these reports.

Outcome

High-accuracy structured extraction at scale, eliminating 50 FTEs of manual work.

Challenge

Financial analysts need reliable answers, but LLMs hallucinate and compliance demands auditability.

Approach

Source-referenced RAG with session memory; prompts shaped by real analyst query patterns.

Outcome

Analysts get answers they can trace back to source documents across multi-turn workflows.

Challenge

Korean regulatory filings in non-Latin script, high volume, tight accuracy requirements.

Approach

End-to-end NLP pipeline with domain-tuned classification and extraction rules.

Outcome

85%+ hit rate across 130K+ docs, 5 FTE reduction, processing cut from 15 days to 40 minutes.

Challenge

Complex financial documents need reasoning, routing, and multi-step validation beyond a single prompt.

Approach

LangGraph multi-agent architecture designed around real document types and analyst workflows.

Outcome

PoC earned regional executive endorsement and demonstrated a path to autonomous document intelligence.

Challenge

A large class of financial documents existed only as images, completely outside automation reach.

Approach

Two-stage pipeline: Azure Layout Model for structure + Python rules for financial semantics.

Outcome

Coverage expanded from 0% to 64%, unlocking a new automation surface and eliminating 5 FTEs.

Challenge

Ingesting financial documents across five markets with inconsistent formats and noisy boundaries.

Approach

Cosine similarity + two-stage classification, with rule-based edge cases handled by domain expertise.

Outcome

3 FTEs saved, 50% reduction in out-of-scope data, reliable ingestion across CA, IN, US, AU.