Boosting AI in Fintech: Analyzing URLs using Semantic Analysis

Boosting AI in Fintech Analyzing URLs using Semantic Analysis

Table of Contents

The fintech sector is rapidly evolving and driven by advancements in AI. As per the Future Market Insights report, the global AI in fintech market is expected to reach USD 58.70 billion by 2034. Thus, we can see improved efficiency, personalized services, and enhanced decision-making in finance. A key impact area is targeted advertising, where semantic analysis is crucial in understanding customer intent and maximizing ad relevance.

This case study explores how High Peak leveraged the power of AI in fintech, specifically semantic analysis and large language models (LLMs), to unlock valuable insights from URLs and enable data-driven decisions about a specific webpage.

About our client

Our client is a leading fintech company in the US, specializing in advertising solutions for financial media. They offer various services, including targeted ad slot placements and content-integrated marketing, to ensure ads are relevant to webpage content. With extensive industry knowledge, in-house technology, and tailored services, they excel at integrating superior content and comprehensive data. Their innovative platform supports publishers and marketers in engaging effectively with financial audiences to drive results.

Key highlights

High Peak developed an AI-powered LLM-based solution capable of accurately classifying web pages based on content type, page type, and the presence of stock ticker symbols using only the URL.

  • We enabled our client’s business value to tailor their ad pricing strategies, leading to increased revenue and improved ROI on ad campaigns.
  • We leveraged advanced natural language processing (NLP) techniques, specifically semantic analysis and cutting-edge large language models (LLMs).
  • We delivered a high-performance solution within 1 month while staying within a strict budget for getting a system ready and a strict budget for the regular execution of the system due to the large volume of URLs requiring analysis (approximately 9 million).

High Peak’s approach to utilizing LLMs for our client’s URL analysis tool showcased our ability to:

  • Rapidly adapt to challenges and pivot strategies to meet budget constraints.
  • Conduct thorough research and testing to select the most suitable LLMs for the specific task.
  • Apply expert prompt engineering techniques to maximize accuracy and minimize hallucinations.
  • Develop a practical and user-friendly API service for seamless integration and client utilization.

Why did the client reach out to High Peak?

The client leveraged our AI development solution to improve the relevance of the ads being displayed on their platform. Their advertising model relies on understanding the content of the pages where ads are displayed. To achieve this, they needed to develop a system using semantic analysis that enabled them to categorize pages and match them with suitable ads. They faced challenges where they needed to:

  • Identify the type of page: Distinguish between news articles, blog posts, landing pages, etc.
  • Classify content category: Determine the subject matter, such as stocks, bonds, cryptocurrency, commodities, or personal finance.
  • Extract financial ticker data: Identify the presence of stock tickers, exchange codes, and the type of financial instrument mentioned.
  • Process a large volume of URLs: Analyze a historical database of 9 million URLs, with the capacity to handle a continuous influx of new URLs daily.
  • Adhere to a strict budget: Find a cost-effective solution that delivers reliable results within a limited budget.

To mitigate these challenges, High Peak built a large language model (LLM)-based solution to analyze URLs for the client. This product aimed to categorize webpages based on three key factors:

  • Page type: Identifying whether a page is a homepage, article page, screener page, or other types.
  • Content type: Determining the content focus, such as bonds, crypto, futures, mutual funds, or other financial topics, particularly relevant in a financial context.
  • Ticker information: Extracting stock symbols (tickers) mentioned in the URL to identify specific stock pages.

This analysis was intended to help the client optimize its ad pricing strategies. By understanding the page type, content type, and ticker information, the company could better target ads and potentially increase their value.

Challenges High Peak faced during the development phase

Despite meticulous planning, our team encountered challenges. These complexities during the development phase illuminate the intricate journey towards delivering innovative solutions. Let’s see how High Peak encountered major roadblocks:

Balancing accuracy and cost-effectiveness.

  • Processing the content of 9 million URLs, plus the ongoing daily influx of new URLs through LLMs, proved to be far too expensive. The client’s budget wouldn’t accommodate this approach, prompting a reassessment of the strategy.

After creating a pricing sheet to model the costs, our AI strategy consulting team realized that relying on content analysis would require two expensive services:

  1. A web scraper: To fetch the content from each URL.
  2. The LLMs themselves: To process large volumes of text.

This dual cost structure would have dramatically inflated the project’s expenses, exceeding the client’s budget by a factor of three or four. This realization prompted High Peak to pivot to a more cost-effective solution: analyzing only the URLs.

  • High Peak recognized that URLs often contain valuable structural and semantic information that can hint at the content of the page.
  • By carefully engineering prompts and training the LLMs on URL structures, they could potentially extract the needed information without relying on full content analysis.

This shift from content analysis to the semantic analysis of URLs marked a crucial turning point in the project. It allowed High Peak to reduce costs significantly while still maintaining a focus on accuracy. The success of this approach ultimately shaped the final design and architecture of the AI-powered URL analysis tool.

Mitigating LLM Hallucination

Relying on URLs rather than full-page content for analysis increased the risk of LLMs generating inaccurate or irrelevant outputs, known as hallucination. This was particularly concerning as the client needed highly reliable information for targeted advertising. High Peak had to develop strategies to minimize hallucination and ensure the accuracy of the extracted information.

Adapting to inconsistent LLM outputs

Early in the development process, High Peak discovered that some LLMs, particularly Llama, produced inconsistent output formats, making it difficult to parse and extract the required information reliably. This inconsistency required additional effort to standardize outputs and ensure compatibility with the overall system design.

Addressing model-specific prompting requirements

High Peak encountered challenges in adapting prompts to the specific requirements of different LLMs. Some models, such as Amazon Titan Text, required specialized tokens and instructions, adding complexity to the prompt engineering process. This meant that our team had to invest time and effort in understanding the nuances of each LLM and tailoring prompts accordingly.

These challenges highlighted the complexities of developing a robust and efficient AI-powered solution. High Peak’s ability to overcome these obstacles demonstrates its expertise in LLM management, prompt engineering, and system optimization.

Development journey

High Peak demonstrated its expertise in semantic analysis and rapid product development by delivering this URL analysis tool within 1 month. Initially planning to scrape website content, High Peak pivoted to a more efficient approach—direct URL analysis using LLMs—after recognizing the high cost of the initial strategy. 

This agile shift, driven by daily collaboration with the client, showcased their responsiveness to budget and time constraints. High Peak’s skilled prompt engineering and rigorous LLM testing resulted in a high-performing solution that accurately identifies page types, content categories, and financial instruments from URLs. Thus, ultimately achieving the client’s goals for targeted advertising.

Product success

  • High accuracy: High Peak took a sample of URLs and achieved a significant accuracy with Claude 3 Haiku and with Mistral 7B in classifying URLs across page type, content category, and financial instrument identification.
  • Cost-effectiveness: The solution successfully addressed the client’s budget concerns. By pivoting away from web scraping and optimizing for token usage, High Peak brought the cost down significantly.  
  • Flexibility and problem-solving: High Peak demonstrated agility by quickly adapting to challenges. The shift from web scraping to direct URL analysis and the refinement of LLM prompts based on client feedback showcase their problem-solving expertise.

These factors suggest that the product we created using semantic analysis successfully met the client’s needs for targeted advertising by providing an accurate, cost-effective, and timely solution.

Technologies used

Programming language: Python 

API Framework: FastAPI 

LLM Hosting Platform: Amazon Bedrock 

LLMs: Claude 3 Haiku, Mistral 7B, Chat GPT 4-o, Chat GPT 4-o mini, Llama 3.2, Amazon Titan Text, Mixtral 8*7B

Leverage High Peak’s mastery in prompt engineering and semantic analysis to empower your fintech solutions

At High Peak, we’ve got the best AI consulting and development team in the industry, ready to tackle any fintech challenge you might face. 

With our expertise in AI services, prompt engineering, and semantic analysis, we invite you to reach out and entrust us to help you with your problems. 

We’re here to ensure they’re resolved, propelling your fintech solutions to new heights with our comprehensive support and innovative strategies. 

Contact us today and get best-in-class AI development services!