Ntropy
Ntropy Technical Briefing
Published May 16, 2023
Supported by IowaEDA and Brale
Overview
Amelia Niemi, co-founder and CTO of Ntropy, walks through the Ntropy transaction enrichment and categorization API. Raw bank transactions go in, structured data comes out — category, merchant identity, logo, location, website, recurrence pattern, active and cancelled subscriptions, and detected income sources. A live Python quickstart shows four lines of code enriching a 1,500-row CSV (roughly 12–18 months of transaction history); the Ntropy app then surfaces the same data in a PFM-style UI. Amelia walks through Ntropy's benchmarks against GPT-4 — about 2,000× cheaper per transaction, 150× faster, and 15× more accurate on the hardest cases — before ML engineer Robin demos "Ntropy Cookie," a Discord chatbot that lets end users ask natural-language questions about their own spending: essentials vs non-essentials, savings opportunities, bank fees.
0:00 Introduction to Ntropy
Amelia Niemi, co-founder and CTO of Ntropy, introduces the Ntropy transaction enrichment and categorization API — raw bank transactions in, structured merchant data out.
1:00 API overview: income, recurrence, categories
The Ntropy API ingests raw bank transactions and returns enriched data — categories, merchant identity, recurrence patterns, income source detection, and all the bells and whistles fintechs need for PFM, underwriting, and personalization.
2:30 Python quickstart — five lines to enrich a CSV
Live quickstart from the Ntropy docs: initialize the SDK with your API key, point it at a CSV of 1,500 transactions (~12-18 months of history), and save the enriched output. Four or five lines of Python.
4:30 Enriched transactions in the app UI
Uploading the same CSV into app.ntropy.com shows labels, merchant logos, locations, website URLs, and merchant IDs — the same API output rendered in a PFM-style interface.
5:30 Subscriptions, recurring payments, and income detection
Out of the box the app surfaces active and inactive subscriptions, recurring payments (including monthly Spotify), and detected income sources like salary — useful primitives for personalization and underwriting.
7:00 Ntropy vs GPT-4 benchmarks
Amelia walks through benchmark numbers for Ntropy's specialized transaction enrichment vs GPT-4: roughly 2,000× cheaper per transaction, 150× faster latency, and 15× more accurate on the hardest cases.
8:30 Ntropy Cookie — chat over your transactions
Robin demos "Ntropy Cookie," a Discord-hosted chatbot built on top of the enrichment API — users ask "how much did I spend on fast food last month?" and an LLM reads enriched Ntropy output to answer.
11:30 Pre-made prompts: essentials, savings opportunities, bank fees
Shipped prompts show essentials-vs-non-essentials breakdowns (correctable by the user), savings opportunities, bank-fee audits, and a tongue-in-cheek "roast me based on my transactions" command.
Topics: Data Infrastructure, Developer Tools
