The foundation of Conversational Finance — the hila Platform

June 28, 2024

Conversational Finance powered by hila provides a rapid way for an enterprise to use generative AI productively. Conversational Finance provides a simple conversational user interface on top of financial databases, such as ERP systems, but that simplicity belies deep technology at its core.  

This technology, built over several years, provides the basis for Conversational Finance, including the ability for the application to be deployed in air-gapped environments or on private clouds. What hila’s Platform brings to Conversational Finance:  

Content not code

A key attribute to the hila Platform is in its extensibility with the ability to customize with no code. The same platform can be adapted for new domains by adding new content, training and knowledge. For example, Legal groups can utilize the various querying capabilities that the Finance team can, such as asking against contracts or performing advanced analysis on many contracts at once.  

This is further enhanced by customization on user-preferences and customer-specific data. This customization extends across all dimensions specific to an enterprise, such as the style of questions asked, the jargon used, and the knowledge specific to the company, such as the fiscal year. For example, one dataset may consider “profit centers” a core part of the business, while others may call them business units or otherwise. In that case, the system can interpret the following question to maintain reliable responses based on the company-specific knowledge.

What was the total amount for the account category Interest Expense in fiscal year 2024 by Profit Center

The SQL this question generates changes based on the knowledge of the company. For example, if the company’s fiscal year runs from April 2023 to March 2024, Conversational Finance will automatically generate SQL to get results for that timeframe.  


We have multiple, patent-pending technologies to eliminate hallucinations in generative AI on both structured and unstructured data. We began work on hallucination work over a year ago, identifying it early as a significant concern with generative AI. Now, these anti-hallucination techniques are agnostic to the content that they exist on.  

In this way, we can ensure this technology, which resides on the platform, works regardless of the domain, department and enterprise. There is far more information about how these processes work in each of the following blogs, on structured and unstructured data.  

For example, our techniques return reliable responses when GPT4 on its own does not return anything. We ran tests that showed that for simple queries, GPT4 responded often, but nearly three out of four times, GPT4 refused to return any data against complex queries.  

More on this table is here.  

Model building and training

Our platform supports model fine-tuning and training, with a very extensible architecture that enables swappable models. This provides the ability to swap out a private, self-hosted model with a publicly available model. This retains all of the other features of the platform, such as the anti-hallucination technologies. This structure also retains the fine-tuning should a new, better open-source model be preferred.  

This methodology can work across various use cases as well. For example, for one client, we fine-tuned a translation model on their behalf, and incorporated it into the overall flow of their application. This fine-tuned model worked in conjunction with other open-source models, as they had a requirement to host our platform entirely air gapped from the internet.  Our fine-tuned translation model achieved BLEU scores 20 percent higher than Google Translate and GPT4 —the leading translator in the industry.

The hila Platform orchestrates between these different models, as it can across various generative AI applications.  

Agentic solutions

This orchestration builds out a capability of agentic solutions that utilize the various capabilities of many models to create an overall application. For example, we have agentic solution that can extract structured data from unstructured documents, such as contracts.  

This too is extensible. By changing just a few parameters, the extraction can occur on a variety of document types for several departments, such as Marketing, Communications, or Procurement.  

We developed this agentic approach as the previous process was completely unusable. Extraction techniques using LLMs on their own, much like the complex queries, completely failed to pull in the right information. After our approach, the accuracy rose to above 90 percent based on the predefined template, and subsequent refinements brought it to nearly perfect.

But this agentic approach is also inside of our RAG processes and various other components. These are simple to use and simple to add to an application, and moreover make the data inside of an enterprise more accessible.  


The hila Platform engages in high-performance monitoring of cost, quality and performance of the LLMs inside of the applications. This accompanies the ability to monitor applications at scale with a robust feature set that includes root-cause analysis, heatmaps, mitigation and validation. A much deeper blog on LLM Monitoring is available here.

This process is accompanied by LLM benchmarks and processes that enable a simpler way to operate LLMs at massive enterprise scale. Our LLMOps can perform on more than 200 billion inferences with sub-second responses. This represents an over 10,000X improvement on the current state of the art.

Cloud agnostic, air gapped deployments

The technology underpinning Conversational Finance enables it to be deployed on premises, air gapped or in a private cloud. This technology provides all of the components mentioned throughout this blog to also exist in any of these environments.  

This means that the hila Platform enables a customer to a privately deploy generative AI instance inside of their environment, with none of their data leaving their landscape, standard LLM monitoring for their admins, and RAG and anti-hallucination processes for their data processing, all of which are part of the solution.  

hila Platform is both robust and designed for ease of use. Years of dedicated engineering have been a foundation for Conversational Finance, so while the system may appear simple for users, it exists on a very complex core.