Endura AI — founder

Multi-Agent Financial Automation

Agentic bookkeeping and reconciliation for chartered accountants

PythonLangGraphDSPyPydanticGeminiRAG

Problem: Month-end close for a chartered accountant is days of moving data between Tally, bank statements, and QuickBooks, categorizing transactions by hand. The work is high-volume, rule-shaped, and unforgiving of errors — a bad categorization is an audit finding, not a bug report.
My role: Founder and sole engineer. Architecture, agents, integrations, and the extraction pipeline.
Approach: Structured extraction over free-form generation. Documents parse through Gemini, but every output is forced through DSPy and Pydantic schemas, so an unparseable record fails loudly instead of arriving plausible and wrong. Retrieval is tenant-isolated at client, company, and folder scope — slower to build than a shared index, but the only version defensible for accounting data.
Outcome: The reconciliation agent ingested 10,000+ bank transactions with confidence-scored auto-categorization, collapsing manual month-end reconciliation from 7–8 days to 3–4 hours.

The domain constraint

Why accounting is a bad fit for confident-sounding LLM output, and what “audit-grade” had to mean in practice.

The client/company/folder scoping — how it’s enforced in the retrieval layer rather than checked afterwards, and what that cost you.

How the reconciliation agent decides what it can auto-categorize and what it escalates. The threshold, and how you picked it.

Optional. Founder projects earn their place partly through what you learned.