solo-developer
Agent automating accounting reconciliation
LangChainLangSmithGPT-4o miniNetSuite ERPSlackGoogle Sheets
“75% workload reduction in AR reconciliation after 2 months development + 1 month data”
solo-developer
Why they built it
Monthly reconciliation chaos (10-12 days, discrepancies, emails); understaffed team 12hr/6days; post-Series C push for AI in finance
What worked
Mini LLMs effective for extraction/matching; human oversight via Slack; 75% workload cut; verifiable junior-level outputs
What broke or was painful
n8n too inflexible for custom logic and human-in-loop; large LLMs had context limits; nano LLMs poor reasoning; generic prompts failed
The result
75% workload reduction in AR reconciliation after 2 months development + 1 month data