Whenever you're ready, walk me through how you'd approach this.
What strategies would you implement to mitigate hallucinations in LLM-generated summaries of earnings call transcripts at The Carlyle Group?
Discuss how you would establish an evaluation baseline, ground summaries in transcript evidence, design prompts and output validation, handle uncertain or missing information, and monitor the system after deployment. Include relevant cost, latency, security, and prompt injection considerations.