Overview
My Role: Product Designer
Impact: Reduced time to explore strategies by 80%
Duration: 6 months
Collaborations:
Arthur Sugden (Mentor)
Isha Raghuvanshi (User Researcher)
Stealth startup (Contract client)
Background
Investment decisions should be rational and data-driven. Yet, young adults struggle to even begin between overwhelming information, expensive services, and complex choices.
Scope
Led the design process including user research, interaction design and creating guidelines for human-AI collaboration in financial advisory.
Problem
Information overload in learning about personal finance
"Beyond learning the basics, I spent hours and hours watching YouTube videos, trying to find information that I could directly apply. I just didn't know who I could trust online and whether the advice applied to my financial situation. It was overwhelming."
— User interview participant
Insights
For young adults who have just begun investing, the financial goal is clear but the path is not.
Information and their sources are hard to judge for trustworthiness and relevance
Execution platforms help to execute but not to strategize decisions and analyze financial health
Trends among investors
Many sources of information
Young adults research at least 6 sources of information .
Social media
Youtube and Reddit are the most popular media for financial knowledge.
Short and long-term goals
Goals include social responsibility and entertainment along with retirement funds.
Opportunity
This is an opportunity for banks to help young investors with credible advice and information by leveraging LLM as an exploration tool
User Trust: Banks are largely perceived as a secure environment backed by regulations and protections
User Data: They have insights and data on the financial history, needs, and habits of an individual
Benchmarked Information and Advice: They are a repository of benchmarked financial information, sources, and advice
Challenge 1
Hallucination and error handling
Human oversight: LLMs hallucinate and make errors. They should be used for exploration instead of decision-making. For high-stakes situations beyond a threshold, human advisors should be looped in to make the final decision after initial brainstorming using the LLM.
LLMs as a synthesizer: LLMs hallucinate. Thus, LLMs should be employed as a text synthesizer rather than a database, crafting concise narratives and patterns from information and sources approved by the bank.
Challenge 2
Guiding investors to structure effective prompts
Guided Prompts: The quality of the answer depends upon the structure of the prompt.
Users should be guided to include intent and context.
System prompt should include parsing through the user's financial data and web search trusted sources for the latest developments and market situations, with consent
Guardrails: The system should prompt the user when it is unsure about an answer for:
Required clarifications, instead of assuming details
High perplexity score where the question is out of scope - is not largely covered in the training dataset
Challenge 3
Structuring answers that investors can trust
Structured Answers: Along with the advice, the answers should contain:
The logic or process by which the advice was arrived at
The information referred to both in terms of user data and the sources
A confidence score in the answer
Further prompts for modification, explanation or next steps
Solution
An LLM-powered Financial Guide in the banking ecosystem to help retail investors explore personalized investment strategies
By integrating LLMs, banks can parse through best practices, financial fundamentals, and market trends, and tailor financial advice to each individual based on their data, helping young adults make financial decisions.
Guided prompts and permissions for better answers
Clarifications and guardrails
Structured answers
Process
User Research: Investment education and LLM perception
Finding a Use-case: User requirement and tech capabilities trade-off
Iterative Design: Low fidelity prototyping and testing
Reflections and Learnings
Drunk Island
Where is performance at scale but low to medium accuracy valuable?
LLM as Summarizer
Understanding strengths: parsing at scale, recognizing patterns.
Human - AI collaboration
2 imperfect entities (Human and AI) coming together to balance each other.
View Other Projects
Opening new stores and expanding branches involves high cost and lead time. I redesigned Datasutram Markets, an ML-driven platform that delivers data-driven location-based insights for business growth to achieve a 75% increase in user retention.
Smart cities are enable by IoT resources for efficient functioning but data collection is opaque. I designed the IoT Privacy Assistant for 17,000+ residents to reduce friction in managing their personal data, reducing the multi-channel process to a single click.
While second-hand furniture is economical, it lacks the emotional connection that helps create a sense of belonging in a new place.
This project explores how shared furniture can help foster a sense of belonging, turning impersonal rooms into their homes.