Manish Sanwal is an AI engineering leader and researcher developing Layered Chain-of-Thought, an approach that checks each stage of an AI system’s reasoning before errors can spread. His work spans language-model evaluation, specialized agent architectures, and the operational safeguards needed to make automated decisions easier to inspect and correct.
His career progressed from software and mobile engineering into enterprise AI. By the 2025 AI Engineer Summit, he was Director of AI at News Corp, where his interests included reasoning, explainability, and automation.
Sanwal’s 2024 research on contrast-set evaluation tested whether language models maintain accuracy when wording changes but meaning does not. An ELECTRA-small model’s accuracy fell from 89.9% to 72.5% on modified inference examples; training with contrast-enhanced data improved performance to 85.5%. His subsequent Layered Chain-of-Thought research extended that concern with robustness into multi-agent systems, combining staged reasoning, external verification, and optional human feedback.
- Intermediate-step verification: Check each inference against external databases, structured knowledge, contextual consistency rules, or independent evaluators before it influences subsequent reasoning.
- Specialized multi-agent systems: Assign distinct responsibilities to modular agents that can be improved independently, with verification between components helping contain failures and support auditing.
- Continuous agent feedback loops: Combine automated graders, human feedback, prompt versioning, monitoring, and rollback mechanisms to keep self-improving agents accountable.
- Conversation-level evaluation: Measure enterprise assistants across the complete user journey, including clarifying questions, task completion, user effort, escalation quality, and handoffs between agents.
