Scenario

You are building an internal finance research copilot for analysts who need fast, grounded answers about issuers, sectors, earnings, and risk factors. The assistant should synthesize internal research notes, earnings-call transcripts, approved market commentary, and policy documents into concise answers with citations. Usage is expected to reach a few thousand analyst queries per day, with heavy spikes around earnings season and market events. Analysts will rely on the tool for draft research and meeting preparation, so unsupported claims are a material risk.

Constraints

• p95 latency must stay under 2,500ms for standard questions
• Cost ceiling is $25K/month at 80K queries/month
• Hallucinated factual claims must stay below 2% on a labeled evaluation set
• Every material claim must be traceable to retrieved sources
• The system must resist prompt injection in retrieved documents and user inputs

Available Resources

• Internal research notes, approved commentary, earnings transcripts, and policy documents
• An approved LLM API and embedding model
• Hybrid search infrastructure with keyword and vector retrieval
• 20 hours/month of analyst labeling capacity for eval sets and adjudication

Question

How would you build this assistant so it delivers useful finance answers while staying grounded in approved sources, and how would you evaluate and operate it under these latency, cost, and safety constraints?