Apoorva Joshi

Apoorva Joshi is a Staff AI Developer Advocate at MongoDB whose work applies lessons from cybersecurity and information retrieval to building dependable generative-AI applications. She teaches developers to match system architecture to actual requirements, using agents only when their added autonomy justifies their cost, latency, and operational complexity.

Joshi studied electrical engineering and completed a master’s degree in computer engineering at the University of Iowa, researching user churn in online health communities. At FireEye, she worked as a research scientist on email security and coauthored research on malicious URL detection that classified suspicious links from their lexical characteristics without crawling potentially dangerous websites.

She subsequently worked in security data science at Elastic, addressing endpoint malware, threat detection, and user behavior. Her security research and writing covered transformer fine-tuning, malware beaconing, and GPT-4-assisted analysis of Linux user sessions. After joining MongoDB in 2024, she advanced from senior to staff AI developer advocate, developing technical education around retrieval, agents, and production system design. Her public GitHub projects include an early comparison of cosine similarity and dot-product retrieval.

• Choose the simplest architecture that works. Joshi distinguishes basic prompting, retrieval-augmented generation, controlled workflows, routing, and autonomous agents according to task requirements. Her production system-design framework applies that judgment to hypothetical insurance-claims review, combining clinical guidelines, coverage policies, and patient history with human-in-the-loop review for complicated cases and potential denials.

• Preserve context across text and images. Her approach to multimodal document retrieval uses page images, multimodal embeddings, vector search, and vision-capable language models to keep charts, tables, and surrounding text together. She distinguishes model-requested tool calls from their execution in application code and identifies page boundaries, metadata, cost, and latency as practical design constraints.

• Rank results by user intent. Joshi uses instruction-following rerankers to incorporate priorities that semantic similarity misses, including evidence quality, troubleshooting intent, recency, and safety-critical details. Her context-engineering guidance and coauthored analysis of specialized rerankers emphasize retrieval quality alongside document-order sensitivity, inference cost, and latency.

• Measure whether the entire system deserves trust. Her production criteria include citation faithfulness, guardrail failures, human overrides, review time, and cost per recommendation—signals that connect model behavior to business outcomes and accountable human decisions.