Anant Srivastava works on data architecture and production AI, helping developers connect the behavior they want from an application to the systems that supply its knowledge. His work spans technical consulting, developer education, vector retrieval, and the architectural choices that determine when to use prompts, external memory, or model training.
From consulting to developer education
Srivastava’s career includes a return to MongoDB after four years, moving from consulting into Developer Relations. That career transition brought his work into developer education: explaining database choices through applications, performance tradeoffs, and practical implementation decisions. At MongoDB, he worked as a Senior Staff Developer Advocate; in 2026, he held the role of Principal Technologist for Data and AI Platforms at Oracle.
His MongoDB teaching connects retrieval-augmented generation to the database underneath it. A comparison of MongoDB Atlas and PostgreSQL with pgvector uses a financial question-answering application to make latency, throughput, and deployment choices concrete. The question is how those choices affect an application’s ability to serve useful answers, rather than which database wins an abstract comparison.
Two connected concerns distinguish Srivastava’s technical work: reducing the cost of retrieval without losing useful results, and recognizing which problems retrieval should solve.
Embedding quantization and rescoring: His work with MongoDB and Voyage AI embeddings examines lower-precision vector representations and their effects on performance and accuracy. Binary quantization can make an initial search cheaper; rescoring then evaluates the selected candidates more precisely. The useful tradeoff is whether the combined process preserves retrieval quality while reducing storage and search costs.
Fine-tuning stable behavior: Srivastava distinguishes the jobs of prompts, memory, and weights. Prompts guide small, stable behaviors. External memory supplies knowledge that changes, requires citations, grows beyond a manageable prompt, or needs access controls. Model weights suit stable learned responses. This distinction makes diagnosis consequential: fine-tuning cannot reliably repair missing or poorly retrieved information, and changing facts need a system that can update and govern them.
These concerns connect his database education to broader AI architecture. Efficient retrieval matters when external knowledge is the right tool; training matters when the desired behavior is stable enough to learn. Srivastava’s work helps developers make that choice deliberately, alongside the practical decisions about accuracy, latency, and cost.
Anant Srivastava explains how to assign behavior to prompts, changing knowledge to memory, and stable task reflexes to weights—and how to move information among them as an agent learns from its work.
Prompt edits, indexed documents and training samples each decide where behavior or knowledge lives. Review their combined effects as architecture.
When an assistant lacks the right document chunks, fix retrieval. Fine-tuning on runbooks can encode stale facts while preserving the original retrieval failure.
Fine-tune stable task reflexes when human judgments have converged and capability or cost justifies the change. Keep humans at contested edges and monitor drift.
▶ Watch ↗