Andrew Dumit

Andrew Dumit is an AI engineer at Watershed developing trustworthy systems for product carbon footprints and supply-chain emissions measurement. His central challenge is that carbon accounting depends on expert judgment: an agent can produce a plausible emissions estimate while making unsupported assumptions, mishandling data, or concealing how it reached its answer.

Dumit graduated from Rice University in 2017 with a statistics degree and worked on modeling a locust’s motion-detecting neuron as a TensorFlow recurrent neural network. He subsequently joined Buoy Health, where he became a senior data science engineer working on medical decision-making. His contributions to federal consultations on explainable artificial intelligence and the NIST AI Risk Management Framework established an early focus on accountable AI in consequential domains.

At Watershed, that focus extends from emissions datasets to agent safeguards and evaluation:

• ATLAS spend-classification benchmark: Dumit co-developed ATLAS, which uses labeled synthetic purchasing data to assess how accurately models connect company expenditures with emissions factors used in Scope 3 accounting.
• Property-based agent evaluation: With Ishaan Parikh, he distinguishes correctness checks from property checks: experts may reasonably disagree about manufacturing assumptions, but incompatible units, impossible material balances, and unsupported claims remain identifiable failures. Their evaluation strategy emphasizes realistic customer inputs, expert feedback, and tests that evolve with the product.
• Governed coding agents: Dumit replaces rigid graph-manipulation tools with coding agents capable of exploring complex supply chains and editing many product models. To prevent untracked changes and falsely reported edits, he routes consequential operations through a typed TypeScript SDK backed by deterministic execution, linting, conflict detection, and validated outputs. The resulting human-readable review artifacts let sustainability specialists inspect changes without reading generated code.

Dumit also co-authored validation criteria for AI-assisted carbon footprinting, extending these concerns to system-level benchmarks, documentation, data quality, and uncertainty.