Sai Krishna Rallabandi

Sai Krishna Rallabandi is director of data science at Fidelity Investments and a language-technologies researcher building privacy-aware shared agents for group conversations and wearable devices. His work addresses a practical problem that conventional personal assistants largely avoid: deciding who is authorized to receive information, where to deliver it, and when to remain silent.

From speech technology to financial AI

Rallabandi began as a research assistant at the International Institute of Information Technology, Hyderabad, before joining Carnegie Mellon University’s Language Technologies Institute, where he worked with Alan W. Black as a master’s research assistant and doctoral researcher. His doctoral research focused on disentangled speech representations, and in 2020 he received a Facebook Fellowship in spoken language processing and audio classification.

His research extended into code-switched language processing and voice-enabled banking for multilingual communities. Public projects including Vocoding-Experiments and Festival-Speech-Synthesis-System document his work with speech generation and established synthesis software.

He joined Fidelity as a senior data scientist in 2022 and became director of data science in 2025. His financial-language research includes length-aware preference optimization for financial question answering and single-agent and multi-agent investment-report generation.

Building assistants that understand a group

Rallabandi’s personal project Judith, which originated during his Carnegie Mellon years, extends across group chats and smart glasses. Its shared-agent architecture centers on three engineering commitments:

• Action-boundary agent security: His Jataayu security layer combines deterministic safeguards and LoRA-adapted small models to separate trusted instructions from untrusted data, intercept prompt injection, and scrutinize sensitive actions before execution. It also accounts for risks created when individually benign tools are combined.
• Selective, context-aware memory: Judith extracts atomic facts, continually reassesses their relevance, and discards information that no longer matters. Retrieval quality, token costs, and key-value-cache behavior shape the resulting memory architecture.
• Role-aware privacy and participation: Shared information is routed according to audience and context, with private responses when public disclosure would be inappropriate. User-specific adapters and lightweight classifiers help determine both who may receive an answer and whether the assistant should respond at all.