An estimated 14.1 percent of U.S. households were underbanked in 2021, as reported by the FDIC, resulting in limited access to financial services such as credit and loans. This lack of access can drive individuals towards riskier financial alternatives like payday loans or check-cashing services. The issue extends beyond the domestic sphere, with many underbanked populations globally lacking basic financial services such as loans or checking accounts.
Recently, I engaged in a discussion with Naré Vardanyan, the CEO of Ntropy, a company that has developed technology capable of categorizing and enhancing financial data on a large scale. Naré established Ntropy a few years back with the aim of tackling these challenges and fostering a more inclusive financial system for all individuals.
Our dialogue covered a wide array of topics, ranging from Naré’s background working for the UN and upbringing in Armenia to the impact of recent advancements in AI and large language models (LLMs) on our comprehension of financial data.
Gary Drenik: What constitutes financial data, and why is its accurate categorization crucial?
Naré Vardanyan: Financial data encompasses the records of capital movements that drive the global economy and our daily financial interactions, encompassing activities such as checking accounts, mortgages, car loans, money transfers, and more. According to a recent survey by Prosper Insights & Analytics, nearly 70% of individuals utilize banking services.
Traditionally, this data has been segregated within individual banks and financial institutions, posing challenges for its effective utilization. The absence of uniformity across companies and accounts complicates processes in financial services, making tasks like risk assessment, compliance, underwriting, and accounting both costly and difficult to scale. Consequently, financial institutions often exhibit low Net Promoter Scores.
By structuring and categorizing this data effectively, all financial workflows powered by this data can be enhanced. Streamlining the data leads to quicker and more cost-effective onboarding, underwriting processes, and personalized service recommendations. Making financial data actionable results in streamlined and scalable financial products and improved decision-making for all stakeholders. This was the foundational premise behind open banking regulations and the subsequent innovations in financial services over the past decade.
Drenik: What are the primary challenges companies face in categorizing financial data?
Vardanyan: In the realm of finance, the absence of a standardized record system means that financial transactions and merchant information often lack consistent naming conventions, resulting in a plethora of unstructured text strings. Deciphering this data requires companies to painstakingly interpret each transaction. Furthermore, the confinement of data within specific models or text blocks within a company hinders its sharing and comprehension across different organizations.
Traditionally, decoding this data has relied on manual efforts or in-house rule-based models, which are time-consuming to construct and typically exhibit an accuracy rate of around 60%. Prior to the emergence of advanced language models, machines struggled to decipher these text blocks, and existing solutions were inefficient, costly, and inadequate for handling such information at scale.
Drenik: How does Ntropy address the challenge of categorizing financial data?
Vardanyan: Ntropy is among the select few companies that are deeply entrenched in the finance sector and native to LLMs. We have structured our entire technological framework around language models from the outset. While not every problem aligns seamlessly with language models due to their sluggishness, costliness, and unpredictability in certain scenarios, they are exceptionally well-suited for tackling challenges involving unstructured text data—such as financial transactions. ChatGPT serves as a prominent example of an LLM, with 65% of individuals familiar with it, as per Prosper Insights & Analytics.
To develop a model capable of comprehending this data, a broad understanding of the world coupled with domain-specific knowledge and feedback loops is essential. Ntropy’s technology amalgamates billions of data points from global merchant databases, search engines, and language models trained on a condensed version of the web to process banking data across multiple continents and languages. Our objective is to distill meaningful insights from diverse transactional information and present them in a coherent, digestible manner.
Drenik: Ntropy leverages AI for classification; how was this technology constructed?
Vardanyan: Transaction classification poses a complex challenge due to the ever-expanding array of intermediaries and merchants, rendering it impractical to address using traditional rule-based methodologies. This is why conventional software approaches have consistently fallen short in categorizing financial data.
For instance, Plaid, a prominent figure in global fintech, has played a pivotal role in democratizing financial data. Despite its efforts to provide organized, categorized transaction data from its inception, it encountered difficulties in doing so efficiently and at scale. The breakthrough in language models has finally enabled us to tackle this issue effectively.
Our journey commenced with a modest training set of 20,000 transactions and a distilled BERT model featuring 400 million parameters (a popular deep-learning model for natural language processing). Presently, Ntropy processes over 100 million transactions monthly, utilizing significantly larger models.
We have meticulously trained our models using raw financial data, expert insights, and machine-labeled samples to ensure unparalleled precision. We consistently share benchmarks to elucidate our methodologies and rationale, fostering transparency.
Today, owing to our expanding user base, we have evolved into one of the most comprehensive entity lookup services globally, encompassing nearly 100 million merchants enriched with detailed attributes like names, addresses, and industry tags. This dynamic database empowers our models to swiftly identify companies and append relevant attributes to financial transactions.
Over the past seven months, our focus has centered on ensuring cost-efficient accuracy. Developing and maintaining an in-house model demands substantial resources. As more financial institutions explore the potential of LLMs for tasks like reconciliation and underwriting, Ntropy offers its services at a fraction of the cost and latency.
Drenik: Your personal background includes experiences in Armenia, the UN, and investments in AI and ML startups. How have these experiences influenced Ntropy’s development?
Vardanyan: My upbringing in Armenia and Armenian heritage have profoundly shaped my character and leadership style. Few individuals in the Western world have encountered conflict zones and war, thereby instilling a sense of urgency that can be a potent asset.
My tenure at the UN and my interest in AI have been instrumental in shaping Ntropy’s trajectory. From the outset, we envisioned Ntropy as a global platform facilitating products and services across diverse organizations, languages, and currencies.
My involvement with Ntropy is deeply personal. As an immigrant and entrepreneur, I have navigated periods without traditional employment. While I have initiated companies and secured substantial funding, I have also encountered challenges in obtaining personal loans and securing apartment leases. Recognizing that a deeper understanding of financial data could address these issues propelled me to embark on this journey.
Drenik: What is your vision for Ntropy’s future?
Vardanyan: Ntropy occupies a distinctive position, boasting extensive domain-specific data and cutting-edge models to automate the back-office functions of financial services. In the near future, tasks that warrant automation—such as onboarding and accounting—will be seamlessly executed.
Why is this significant? Reduced operational costs in financial services can foster mutually beneficial economics for financial institutions, businesses, and consumers alike.
I envision Ntropy as a pioneering platform driving AI adoption and LLM integration in enterprises. We aim to progress methodically, addressing use cases and data types in a scalable manner. By eliminating cost, latency, and reliability hurdles, we can unlock new possibilities.
Operating a bank entails managing systems and teams tasked with risk oversight. Risk management remains a predominantly human-driven and exorbitantly expensive domain. However, by rendering risk management remarkably cost-effective, we can catalyze increased competition and reshape the economics of the system. Consequently, financial services become more accessible and efficient for all individuals—a future we are diligently crafting at Ntropy.
Drenik: Thank you, Naré, for sharing your insights. I eagerly anticipate witnessing Ntropy’s continued growth and its positive impact on various facets of finance.
