Mani Khanuja

Mani Khanuja is a Principal Generative AI Solutions Architect at Amazon Web Services, specializing in the data architecture, security, evaluation, and operating discipline required to put enterprise AI into production. Her work spans machine learning, high-performance computing, retrieval-augmented generation, and increasingly autonomous AI agents.

Before joining AWS as a solutions architect in 2019, Khanuja worked in software engineering, automated testing, and engineering management, including roles associated with SSP and Pitney Bowes. At AWS, she contributed to applications ranging from protein-language-model classification to hierarchical forecasting.

In 2022, she coauthored Applied Machine Learning and High-Performance Computing on AWS with Farooq Sabir, Shreyas Subramanian, and Trenton Potgieter. She subsequently developed approaches to secure enterprise retrieval systems and contextual chatbots grounded in organizational knowledge, addressing how proprietary information can inform model responses without compromising permissions or privacy.

• Enterprise data is the real differentiator. Organizations using similar foundation models produce meaningfully different applications through the quality, freshness, structure, and permissions of their own information. A travel assistant needs customer context and refund policies while protecting personal data; an employee chatbot must respect existing access controls. Khanuja develops these distinctions in her AI Engineer World’s Fair session.

• Retrieval quality precedes answer quality. Her work on intelligent metadata filtering translates natural-language questions into structured retrieval constraints, reducing irrelevant context. She also applies hybrid search, reranking, query decomposition, and incremental data updates, and coauthored research into agentic keyword search that examines retrieval without requiring vector databases.

• Observability makes AI systems governable. Khanuja emphasizes logging retrieval behavior and model responses, measuring context relevance, refreshing stale information, and rerunning automated evaluations after system changes. She treats accuracy, latency, and cost as linked production constraints; semantic caching can reduce repeated inference and response times.

• Security belongs inside the architecture. Useful enterprise assistants must preserve existing permissions, protect sensitive information, ground responses in appropriate material, and enforce policy throughout the workflow. Khanuja incorporates controls such as Amazon Bedrock Guardrails alongside retrieval infrastructure instead of treating safeguards as a separate deployment milestone.