ML System Design at Tochka

Выводы и как готовиться

• The case is a strong regulated-domain ML System Design example: the answer must optimize evidence, auditability and risk, not only model accuracy.
• A local LLM is explicitly constrained by CPU throughput and sensitive data, so the mature design routes only hard fragments to expensive parsing.
• The best system-design thread is blacklist-first candidate generation followed by high-precision context validation and turnover-share aggregation.
• Monitoring needs bank/template breakdowns and reconciliation checks because one bank changing its statement format can break the parser quietly.