Case Study: A Research Assistant for Aster Wealth
Explore the challenges of designing a research assistant for Aster Wealth as a Forward Deployed Engineer. Learn to balance advisor needs and compliance rules, manage conflicting data sources, and apply retrieval-augmented generation to enforce access controls. Understand how to defend solutions addressing real-world constraints in AI system design.
This case stresses the technical round on LLM systems, where “we’d use RAG” is where the questions start. Work through it as if an interviewer were following each decision down to its mechanism.
During discovery, the FDE showed two groups at Aster Wealth the same sample answer to the same question:
Question: As of June 10, what is the maximum fee for this product?
Answer: “The fee is 4%.” Source: report R1, version 2.
The advisor found the answer useful: it provided a specific figure within seconds and linked to the current report. However, the compliance reviewer rejected the answer. The report stated that the maximum fee was 4%, but the answer presented 4% as the actual fee. A second approved report, R2, specified a different maximum of 5%.
The citation pointed to a genuine source, but that source did not support the stated fee. The answer also failed to acknowledge the conflicting information in R2.
Both reviews assessed valid requirements: the advisor needed a timely, usable answer, while compliance required accurate claims and visible source conflicts. The system must meet both requirements.
What is Aster asking for?
Aster Wealth is a US wealth management firm whose advisors are registered representatives of its FINRA-member broker-dealer. To prepare for client discussions, they search an internal research library, compare information across reports, and prepare research notes. Aster wants advisors to find relevant research faster and receive summaries with references they can open and verify. Each summary is intended for advisor review before any information is used in client communications.
Aster had previously reviewed 16 manually selected research notes. In five of those notes, the advisor quoted a report without identifying its version or publication date. This reveals a source-provenance gap in the existing workflow, but it does not establish that the quotations or conclusions were incorrect. Because the notes were selected manually, the findings cannot reliably establish a firm-wide rate.
The next step is to reconcile the quotations against Aster's source records. An assistant cannot establish the source version of an undocumented quotation simply ...