retirement fund

Building AI Assistance for Retirement Financial Planners

Business domain
Finance
Technology
AILangChainChromaDBAzure OpenAI

Background

Our US client, a leader in retirement planning, uses a bucket strategy to balance their customers’ savings. Accumulated funds are divided into segments (buckets) to cover living expenses over different time periods (5 to 7 years, 15 to 25 years, etc.). While short-term segments address a retiree’s immediate needs, long-term investments allow for the growth of the assets.

To support this approach, AltexSoft built and continues to maintain a dedicated software solution for financial advisors. Recently, we enhanced the platform with two AI-powered features—a copilot to help users better navigate the interface and a voice controller for hands-free operation. 

Challenges

Running flexible strategies that require constant manual adjustments can be complicated even for seasoned professionals, not to mention newcomers. Such complexity affects performance and induces additional training costs. The AI copilot and controller, available 24/7, dramatically simplify workflows and address a range of business challenges. 

Speed up onboarding without the need for human intervention

Cut off expenses on education and support

Improve advisors’ productivity

Build up the engagement of existing users and newcomers

Value Delivered

Exploring the system via text and voice

Exploring the system via text and voice

The AI copilot, driven by a large language model, offers an intuitive, conversational way to explore the system, replacing traditional tutorial videos. Whenever facing an issue, an advisor enters questions into a chat interface or just voices them aloud. The assistant responds with explanations and detailed, step-by-step text instructions that are voiced back if the user prefers speaking to typing. The instructions are accompanied by a live demo — a cursor on the screen shows how to complete a particular task.

Addressing the problem of LLM hallucination

Addressing the problem of LLM hallucination

The copilot’s ability to provide accurate answers is significantly enhanced by Retrieval Augmented Generation (RAG), a technology that employs semantic similarity scoring to better understand the intent behind user queries. This approach helps reduce the risk of LLM hallucinations, ensuring that responses are contextually relevant and directly address the advisor's needs.

Analyzing chat interactions to drive continuous improvement

Analyzing chat interactions to drive continuous improvement

The AI copilot tracks the most frequently asked questions and evaluates both successful and unsuccessful responses. This data is used to generate weekly PDF reports, which pinpoint areas for improvement. By continuously analyzing chat interactions, this feedback loop helps fine-tune the copilot’s responses and enhance its overall performance.

Enabling hands-free operations

Enabling hands-free operations

Along with the onboarding assistant, our team introduced a voice controller to facilitate hands-free operations on some pages. This feature allows advisors to interact with the system and perform a range of tasks through voice commands. Users can request an action, such as filling out cells with certain values, and the system will carry out the task on their behalf.

Approach and Technical Info

AI assistance took two months to develop, involving 2 data scientists and a team lead. Improvements and updates followed for another two months.

The tech stack includes Azure OpenAI, ChromaDB (for RAG and semantic similarity), LangChain,  Pydantic, and FastAPI.

AILangChainChromaDBAzure OpenAI