I'm a security engineer at Pinterest and a computer science master's student at Georgia Tech, specializing in artificial intelligence.
At Pinterest, I work on cloud governance, Terraform automation, employee access tooling, and centralized audit logging across AWS and GCP. I collaborate with vendors and internal engineering teams to integrate security tooling and roll out infrastructure security controls.
I'm interested in the safety and reliability of AI systems. My work at Georgia Tech includes agent evaluation and interactive 3D reconstruction visualization; my robotics experience includes policy training and evaluation in simulation.
My current research direction is paired agent evaluation: measuring wrong decisions separately from claim grounding, and checking whether useful context is preserved. The implemented context-integrity extension has synthetic fixtures and a scorer; matched live-model experiments remain pending.
- SafeLattice: Information-flow scoring for language model agent traces, developed as an extension of Claw-Eval.
- Never Trust the Context: A policy prototype for controlling how language model agents use untrusted context, with offline evaluation fixtures and regression tests.
- Grasp Recovery under Delayed Feedback: Implemented a MuJoCo evaluation of grasp-recovery interventions under stale observations and response delay, with 400 pilot episodes and 88,728 matched branches. Preliminary simulation results show no completion advantage from latency-aware remedy selection over the latency-agnostic baseline.
- Grasp Failure and Recovery: MuJoCo grasping infrastructure and HUG integration diagnostics. Failure prediction and recovery experiments under changing mass and friction are in development.
- LeRobot Behavior Cloning Evaluation: ACT and Diffusion Policy training, checkpoint comparisons, and simulation evaluation.
- GTSfM Visualization: Interactive 3D reconstruction visualization, including hierarchical merge playback.
- Franka IK and Viser: Damped least-squares inverse kinematics, joint-limit handling, and interactive visualization for a Franka arm.
These projects are in progress, using original synthetic fixtures, deterministic baselines and small CPU checks. Independent validation and real-model studies remain future work.
- Chip Memory Lab: Prefix-cache experiments, capacity accounting and analytic data-movement scenarios, with a CPU toy-attention correctness check.
- Culture Data Lab: Annotation disagreement and chronological forecast baselines for synthetic creative-content feedback.
- Narrative Context Lab: Controlled original stories testing framing sensitivity, explicit-rule compliance and evidence provenance.
- Health Context Lab: Synthetic research-record corrections, unresolved evidence and action authorization; domain review pending.
- Pinterest — Software Engineer, Security Platform (October 2025–present): Cloud governance, Terraform automation, access tooling, and audit logging across AWS and GCP.
- GEICO — Software Engineer (May 2024–October 2025): Infrastructure APIs and server-inventory data pipelines using React, GraphQL, PostgreSQL, and Python.
- Palantir — Software Engineering Fellow (December 2023–January 2024): A fellowship focused on backend services, containerization, and deployment workflows.
- Comcast — Software Engineering Intern, Applied AI (May–August 2023): CLIP-based content moderation and retrieval using Flask, React, Elasticsearch, Docker, and Kubernetes.
I coauthored Hierarchical Structure-from-Motion Scales FeedForward Reconstruction, presented at the ECCV 2026 SfM-DL workshop. My contribution focused on interactive reconstruction visualization and supporting software.