About the Project:

Wealth of High Net worth Individuals has been historically managed by humans. HNIs, like any sane investor, do not place all their eggs in one basket. Their assets are distributed across classes like Currency, Bonds, Insurance, Stocks, Pension Funds, etc. Wealth Managers typically summarize their portfolio at a weekly or monthly cadence.

However, what happens when an HNI needs information regarding their portfolio, and they need it immediately or else they might lose the opportunity for some fantastic gains. In this scenario, the Portfolio manager will dig up information from several places that might not be consolidated on a dashboard if the Fund Manager is not that established. Even if it is, she must scour for it by applying filters, aggregating, and then summarizing her findings for her client. This is a time consuming process if you aren’t a veteran analyst. Even if you are, this model cannot be scaled as it is expensive to hire and train able analysts. Hence, automation is needed.

‍

Approach:

We intended to replicate the process of a human providing information to their client, however, behind the scenes, it would be a powerful AI running the show instead of a human. The AI should not only be able to contextualize and understand the queries asked by the human, but should be able to accurately fetch the requested information and provide the answer in natural language, just like a human would.

This is how the flow of data looked like for us. The main challenges were that data was scattered across places, and different data sources could point toward different answers for the same query. Hence, it was important to establish which data source to refer to, for what kind of query.

To do this, we built a multiplexor that was able to prioritize the data sources according to the type of query asked.

The next challenge was to make the scattered data sources readable for the AI. SQL databases can be read only with SQL queries, Documents can be read only with OCR, and Website content must be scraped before it becomes readable.

We built APIs whose query parameters were filled using LLMs. These APIs formed SQL queries from Natural Language and queried the required information from the SQL database. Textract helped in extracting information from PDFs and images, while custom built scrapers helped in getting it from websites.

Once the correct information is extracted and understood by the LLM, it was easy for the LLM to provide the response in natural language, in the style in which the question was asked.

‍