CV

B.S. in Computer Science

Georgia Institute of Technology

• GPA: 4.00

Member of Technical Staff

Decagon

• Enhanced an opaque simulation framework for CX agents through LLM-based diagnosis of chat logs against a data-driven failure taxonomy, iterating with real users to create truly useful automations that accelerated debugging by 3x across 100+ deals.
• Integrated customer telemetry into Decagon’s observability platform, unblocking agent hill-climbing over customer-indexed success metrics, which uncovered improvement opportunities that closed multi-million dollar contracts.
• Consolidated alerting into a standard framework that eliminates coverage gaps and reliably detects incidents.
• Built a system to enhance agents by translating customer test specs into formal simulations, identifying misalignment between programmed and anticipated agent behavior, and iteratively co-optimizing agents and simulations.

Co-founder and Lead Researcher

Trace AI Labs

• Created PACT, a multi-metric evaluation framework to assess whether enterprise AI assistants follow compliance rules in 48 realistic scenarios across regulated domains like healthcare, revealing no existing model is reliably safe (preprint under review).
• Analyzed the psychology of why AI agents violate legal constraints under user pressure tactics, showing alignment between LLM behavior and prevailing human behavioral theories like deterrence and legitimacy (AIES 2026).

Research Assistant

Georgia Tech — Human-Centered AI Lab

• Advised by Dr. Mark Riedl.
• Instrumented an explainable system that profiles LLMs’ strengths and weaknesses and selects the best model for tasks under cost constraints (MLSys YPS 2025), and generalized it to routing complex agent workflows (HCXAI at CHI 2026).
• Traced where social reasoning in LLMs originates in pretraining corpora using training-data attribution (COLM 2026).
• Developed counterfactual explanation methods that translate complex agentic-workflow behavior and errors into clear, actionable insights for developers through a visual, interactive interface (HCXAI at CHI 2026).
• Led FLaME, the first holistic benchmark of LLMs over diverse financial analysis and reasoning tasks (ACL Findings 2025).

Research Assistant

Georgia Tech Research Institute (GTRI)

• Built an agentic sandbox with a chat interface for creating, executing, and debugging scientific experiments, infusing context from retrieved research papers and code to empower researchers to rapidly iterate on their ideas.
• Proposed novel graph algorithms for decomposing scientific claims into atomic parts for automatic verification.
• Architected MethodBench, a framework for evaluating LLM competence in determining the solution to open-ended, but previously solved, scientific problems (under review; DARPA Contract HR001125C0302).

Software Engineer Intern

Two Sigma Investments

• Designed metrics to provide traders actionable feedback on their stock ideas and market timing, surfacing suboptimal trading patterns and informing more profitable quantitative models distilled from human trading intuition.
• Built an AWS-native ETL pipeline using Lambda and Step Functions to process metrics in Postgres, incorporating observability, testing, and documentation to ensure reliable and rapid insight delivery to quantitative researchers.

PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?

Under review

Okamoto, M., & Erol, A. K.

Why Do AI Agents Break Rules? How Framing, Context, and Social Signals Shape Compliance

AIES 2026 (also at the COLM Workshop on Agent Behavior)

Okamoto, M., Erol, A. K., & Erol, K.

Where Does Social Reasoning Come From? Capability Provenance in Language Models

Conference on Language Modeling (COLM) 2026

Matlin, G., Chakraborty, C., Eom, S., Okamoto, M., et al.

Encoded Early, Used Late: Where Transformers Begin to Act on an Inferred Partner's Expertise

COLM Workshop on Scientific Understanding of Foundation Models 2026

Okamoto, M., & Sarti, G.

Explainable Model Routing for Agentic Workflows

HCXAI Workshop at CHI 2026

Okamoto, M., Erol, A., & Riedl, M.

Method Bench: Evaluating Path-to-Feasibility Reasoning in the Sciences

Under review

Johnson, B., Okamoto, M., Kerce, J. C., & Fekri, F.

Counterfactual Explanations for Agentic Workflows

HCXAI Workshop at CHI 2026

Singh†, M., Kim†, G., Okamoto†, M., et al. († equal contribution)

DeepVerify: Evidence-Based Expert-Level Scientific Claim Verification

Preprint

Xiong, S., Gungordu, O., Johnson, B., Okamoto, M., Kerce, J. C., & Fekri, F.

FLaME: Holistic Finance Language Model Evaluation

ACL Findings 2025

Matlin†, G., Okamoto†, M., Pardawala†, H., Yang, Y., & Chava, S. († equal contribution)

Trust by Design: Skill Profiles for Transparent, Cost-Aware LLM Routing

MLSys Young Professionals Symposium (YPS) 2025

Okamoto, M., Erol, A., & Matlin, G.

Enhancing Military Family Readiness and Resilience Programs Through LLM-Generated Synthetic Data and Natural Language Interfaces

Military Health System Research Symposium (MHSRS) 2025

Iyer†, A., Okamoto†, M., et al. († equal contribution)

Genome-Wide Association Study of Age-Related Hearing Loss in CFW Mice Identifies Multiple Genes and Loci, Including Prkag2

Journal of the Association for Research in Otolaryngology (JARO)

Polesskaya, O., Boussaty, E., Cheng, R., et al.

RATTACA: Genetic Predictions in Heterogeneous Stock Rats Offer a New Tool for Genetic Correlation and Experimental Design

bioRxiv

Johnson†, B. B., Sanches†, T. M., Okamoto†, M., et al. († equal contribution)