Accounting agents

Tools and context

I designed the multi-agent copilot with LangChain and LangGraph, hybrid retrieval, accounting tools and MCP integrations.

Learning from failures

I built evaluations and Langfuse tracing to inspect answer quality, citations, latency, cost and failure modes.

Changes to contextual email generation reduced quality regressions by 68% and critical errors by 89% in internal evaluations.

In a separate agent evaluation, the selected configuration reduced latency by 32% and cost by 9%, with better grounding and citations.

Ingestion at scale

Temporal workers, bounded concurrency and batched vector writes made document ingestion reliable at approximately 340,000 workflows per week.