Understanding your Bank Statement in 100ms
summary
Claim: Ntropy parses financial transactions in under 100ms with near human accuracy. Pipeline details: attention‑based NER with multilingual training, generalized cross entropy loss and smoothing to resist noisy labels, cluster labeling and augmentation, a cross-encoder re-ranking step hitting search engines for merchant normalization, and multimodal multinask models for higher‑level categorization. Scale and impact: over 1 billion daily transactions, clustering and testing pipelines, production thresholds and regressions guard accuracy, and examples show a thousandfold speedup from ~100s to ~100ms.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:29Bank transaction data complexity
- 4:41From 100 seconds to 100 milliseconds with ML
- 6:46Entity recognition in transactions
- 11:19Multilingual entity models
- 17:37Model stability and explainability
- 21:31Normalizing merchant names
- 28:57Transaction categorization
- 32:08Transaction understanding challenges
- 35:59Q&A
related talks
39mHerding Entities- Information Search and Synthesis in the Context of Transaction Data
24mRetail Transactions at scale for AI - From PoC to production
14mCloud-Native AI for Fraud Prevention with Bayesian Networks
33mSmart Banking - Real Time Driven
33mTrillion-Token Pretraining
21mTransforming Legacy Banking Systems through Machine Learning
41mMining Noisy Transaction Data with Neural Nets
19mFrom Data to Dollars: Improving Medical Billing Accuracy Using NLP