I Built a Bank Analytics Pipeline with LLM Agents: Where It Blocked 7 out of 31 Runs
OpenAI
Anthropic
The author describes a bank analytics pipeline that uses open LLMs to generate research on banks from public data streams. It operates a public site (ortant.net) with 24 published studies and 10 bank dossiers. Of 31 runs, 7 were blocked by a publication gate, each with a reason code. The article details the barriers, costs, and lessons learned.
The author built a bank analytics pipeline using open-weights LLMs to analyze public data. The system runs on ortant.net, offering 24 accepted studies, 10 bank dossiers, and a custom research builder. Each run processes seven public data streams through six AI roles: five stream reviewers, a research agent, three advisors, a chairperson, a claim verifier, and a reporter. The median cost per run is $0.132 and it takes 12.8 minutes. Of 31 runs attempted, 7 were blocked by a publication gate: four failed synthesis, two contained prohibited lexicon, and one had a false positive on personal data detection. The system includes four barriers from data passport to publication gate. The author also notes that a predictive layer was worse than a naive baseline and was removed, and that model advice improves coverage but not accuracy. The article emphasizes transparency, with costs and methods openly displayed.
- Abbreviations
- LLM = Large Language Model — Большая языковая модель
Source: Habr — хаб ИИ —
original