Mohak Sharma

Mohak Sharma is co-founder and chief executive of HoneyHive, a platform for evaluating and monitoring AI applications. He focuses on a stubborn production problem: conventional benchmarks can make agents and retrieval systems look reliable even when they fail actual users.

Sharma studied operations research at Columbia University, where he met future co-founder Dhruv Singh. After working in quantitative trading at Citi, he joined Templafy, leading work on its data and machine-learning platforms. He and Singh founded HoneyHive in late 2022; the company launched publicly in April 2025 with $7.4 million in funding, including a seed round led by Insight Partners.

What makes AI evaluations useful

• Evaluation-driven development: Sharma treats evaluation as an ongoing development discipline. His MongoDB and HoneyHive guide measures retrieval and generation independently so engineers can identify whether failures originate in search, selected context, or generated answers.

• Production-to-evaluation feedback loops: Real customer failures should become labeled regression tests. Sharma argues that developer-written examples quickly lose value when production introduces ambiguous requests, unfamiliar combinations, and shifting business requirements.

• Domain-expert-calibrated LLM judges: Generic evaluation prompts often reward characteristics customers do not value. Sharma advocates having specialists define quality, critique automated judgments, and refine grading prompts with concrete examples; F1 scores or correlation measures can then track agreement between automated and human assessments.

• Criteria drift and dataset drift: Evaluations fail when scoring standards diverge from business needs or test cases stop resembling real traffic.

• OpenTelemetry-native AI observability: HoneyHive’s Python and TypeScript instrumentation connects distributed traces with feedback and dataset curation, while its offline evaluation workflows test prompts, models, retrieval pipelines, and tool use. Sharma has also emphasized agent-trace visualization, annotation queues, experiment dashboards, and alerts: practical infrastructure for turning observed failures into measurable improvements.