A large financial services firm analyzes over a billion transactions every day. Each of these has close to 1,000 features, and the features and the transactions are all subject to seasonality. This combined to have a massive scale challenge for the company.
The company needed to monitor drift of all transactions and accompanying features over set time periods — such as quarter over quarter, year over year or week over week — analyzing the data to rapidly understand and fix any issues in the data for their large algorithm. They also needed the ability to do hotspot analysis, and jumping into the data in a specific timeframe, such as the same week from one year to the next.
hila Enterprise enables our client to analyze massive amounts of data at very low cost. For the financial services firm, there could be hundreds of trillions of datapoints, yet using a mixture of incremental data aggregation and proprietary techniques for analysis on-the-fly, hila Enterprise enabled our client to thoroughly analyze all the data. They also set up alerts for drift and feature monitoring and have a flexible method to set policies. They also can apply this monitoring to LLMs or different model types.
The analysis on more than 200 billion inferences occurs with sub-second responses. There are no large clusters required, and the speed up for a policy run, based on 1 billion records per day, is 10,000x competitors.
Sub-second Responses
0 Required
10,000x
Analysts at a large bank in southeast Asia provide large commercial loans to various companies around the world. Often, these loans require thorough analysis and rigorous understanding of new areas of knowledge — these could include mining in a new country or sales of a certain good in a new market. The team that provides these loans has to respond quickly with depth to ensure the quality of the loans remains high.
The variety and size of the data often requires rapid information intake, and the amount of risk inherent in the loan depends on the bank’s knowledge of the industry and the circumstances of the country. Increasing this knowledge often is time-consuming and tedious or would require expensive outside consultants. Moreover, the bank works privately — all data must be kept within its firewall, and natively in a language not offered by many LLMs and requires language flexibility with English and non-English sources.
We built translation services that outperform Google Translate and work within their environment. And, then we used a mixture of public and fine-tuned models to ensure that the public information they bring in can mix with their proprietary information, without any risk for questions or leakage to the outside. The combination is a holistic, non-hallucinating research application for structured and unstructured data that rapidly enables researchers to understand new topics in foreign countries.
Our anti-hallucination methodology identified sentences to be 60 percent more hallucinatory, eliminated them all, and improved the answer from factual sources. Our fine-tuned translation model doubled the accuracy of the base models and provided a BLEU score 5 points over industry-leading Google Translate and nearly 10 points over GPT4.
20+%
100%
speed & accuracy
A multinational energy company wanted to bring generative AI to contract analysis — using Large Language Models to do a “first pass” on the thousands of documents they have on hand. This analysis varied from simple metadata extraction to complex assignment of risk based on a set number of parameters.
The legal department had thousands of contracts, all with a hierarchy, which they needed to keep organized, while having a simple, intuitive search to look through them on the fly. This was true of past contracts as well as potentially understanding future contracts. The analysis was often time consuming, and if not done, there could be risks held in legal agreements that could go unnoticed — often due to previous legal analysts with knowledge of the agreements leaving.
We developed an agentic approach to extract metadata and do advanced analysis on the contracts. This approach uses multiple LLMs working in concert with each other and, essentially, pulls out the relevant information, does a “first pass” with legal expertise, then organizes the information in a way that’s easily accessed in the future. This approach ensures we get the best of each process and costs 4X less than LLMs with large context windows.
Our system analyzes thousands of contracts, extracts dozens of key legal terms, eliminates hallucinations and completes a “first pass” for legal agreements, previously done by junior lawyers or outsourced labor. It does all of this within a half hour, whereas the manual processes of the past could take months of labor by dozens of employees.
1000s
under 30 mins
legal terms
A multinational CPG company had 42,000+ dashboards that tracked a number of its supply chain, sales and operations functions. To receive an update, the executive team would need to request a dashboard from an ERP, through multiple internal bureaucratic layers, then wait for many weeks before getting a response.
The complexity of the systems of record required a specialist with access to the systems to first bring in the portion of data that served them best. The executives may have simple questions and would like to see a chart with the information, even in the meeting, in real time, but didn’t have access to the system nor the knowledge to make it work — frequently SQL and Tableau.
We built a kinetic dashboard that used our text2sql pipeline, with the guardrails to eliminate hallucinations and generate charts on the fly. This combination enabled any executive or person with access to the database to receive back a chart and table back from a simple, natural language query. The work done with this CPG company is reflected in our Conversational Finance application, which is specifically tuned to an ERP system and has been further enhanced with additional anti-hallucination features.
Our work resulted in improving the text2sql generation accuracy by 96 percent, and with our other services increased that to 100. We also helped reduce the time to create a chart to less than a minute and aided in the company’s overall reduction in dashboards. Finally, we did all of this while keeping the client’s data completely private and inside of their environment.
42,000
from months to seconds
data security
One of the largest pharmaceutical manufacturers in the world regularly uses financial data from its ERP system for planning, strategic decision making, headcount allocation, cross-functional analysis, investor relations and the like.
The pharma company relies on analysts to maintain dashboards connected to their ERP system. These dashboards are many, require weeks to build and often are out of date the moment after the data is pulled. The result is a “good enough” analysis and set of information that key stakeholders use for significant decisions.
Conversational Finance, powered by hila Enterprise, provides various individuals inside of the pharma company with real-time charts and analysis on their ERP system, done without any need for Tableau, SQL or any specialized knowledge. Working with our partners, we had a system emplaced and delivering value in less than a quarter.
The company now benefits from having a real-time mechanism for making data-based decisions without the need for sending a call off for a specialist. This has reduced the many dashboards our clients have and improved response times from weeks to seconds for a question.
30 seconds
0 Required
96% +