Experience

Role: Systems and Infra SWE PhD Intern
Location: Menlo Park, CA, USA
Duration: May 2026 – Aug 2026
Team: Similarity Understanding (under Facebook Content Creator)

• Worked on capacity isolation for the Originality Signals Platform (OSP), which is responsible for originality scoring of Facebook content (images, Reels, and Stories). The originality models share the same predictor resources (CPU, RAM, and GPU) without per-model isolation or visibility, so a usage spike in one model could impact the other models and the whole platform.
• Designed an isolation system applying Meta’s Resource Isolation and Management (RIM) framework. Automated the modeling of per-model resource usage across the three dimensions and registered the models as same-priority tenants under RIM’s GlobalFairShare policy. Models can use idle capacity when needed, and under resource pressure the least-priority tenant is shed first (the highest-traffic one among same-priority tenants), leaving the other models and the platform unaffected.
• Onboarded 10 originality models (nine video and non-video OCUC models plus the content cloner model) onto the isolation framework, with per-model visibility dashboards showing resource usage trends and alerts for critical issues such as shedding.
• Built an automated calibration pipeline in which a scheduled Chronos job finds a reliable linear regression fit per model across the resource dimensions, and an async job runs a bot that creates a code diff with the updated weights, which auto-lands after human review.
• Developed a retry pipeline on Dataswarm that retries failed model requests for up to 3 consecutive days before moving them to a dead-letter queue.
• Extended the isolation design to run backfill as a low-priority sheddable tenant that only uses idle capacity. Together with the isolation infrastructure, this is projected to save more than $500K per year, given roughly 5 billion backfill requests per model each month, and free up capacity to onboard new models under OSP.
• Developed Claude skills that automate onboarding a new model, running calibration, and adding a new resource dimension, and followed an AI-native workflow with SecondBrain and Claude skills to design, implement, and review the work.

Role: Systems and Infra SWE PhD Intern
Location: Seattle, WA, USA
Duration: May 2025 – Aug 2025
Team: Data Flows (under Privacy Org.)

• Designed and implemented an AI agent that explains why and how sensitive data flows between source and sink assets, a gap left by Meta’s runtime analysis tool (Probes), which detects flows but not their cause. This cut expert review time from ~20 minutes to under 2 minutes per evidence.
• Built an iterative reasoning pipeline over read/write stack traces that pinpoints the exact code location of declassification, or justifies a preserved data flow with supporting code.
• Achieved ~94% accuracy in identifying declassification at the source–sink pair level (~70% precision/recall per evidence, ~78% on complete flows), evaluated against static analysis and admin-approved ground truth.
• Shipped the agent into the reviewer UI with its full reasoning output, plus monitoring and cost alerts; extended a cross-team multi-agent AI framework for this use case.
• Designed a lineage graph visualization on a fast graph-traversal API and contributed new features back to the API.

Role: Software Engineering Intern
Location: New York City, NY, USA
Duration: May 2022 – Aug 2022
Team: Privacy Approval Monitor (under Messenger Org.)

• Developed a UI tool that detects sensitive database access in code blocks, used by team leads to vet Messenger features before production release.
• Developed a second UI tool visualizing data flows across Meta’s privacy assets (e.g., mailbox API → stored procedure → database), working with cross-functional teams to onboard the target datasets into Meta’s lineage system.

Role: Software Engineer
Location: Pune, Maharashtra, India
Duration: Jul 2019 – Aug 2021
Team: Global Credit Module (GCM) (under Wealth and Personal Banking (WPB))

• Full-stack developer on the credit product used daily by relationship managers across HSBC’s wealth banking.
• Delivered customer-facing automations including email notifications for credit limit approvals and rejections, SMS alerts for credit margin status, Jasper reports, and rule assignment for securities received from batch processes.
• Monitored production batch runs and managed code during production releases.