AI org, Glean Assistant (Agents + Chat) ML Team. Applying LLMs to build "Work AI for all". Projects include:

• Agentic File System (AFS): Built + shipped AFS system for Glean Chat: in-house cloud-based agent sandboxes with shell tool access to file system, programmatic tool calling, and skill management, to drastically accelerate co-working product offering.

• Code Interpreter/Writer: Took full ownership to drive performance + quality on data/file analysis queries in Glean Chat: built + shipped new agentic reasoning/looping system (PyAgents) atop OpenAI Code Interpreter APIs + Glean Context/Indexing, reducing latency by 40% and user-facing errors by 70%. Also optimized UX (with PM + Design collaboration).

• Query Classification (QC): Designed + built + shipped E2E LLM-based QC system to run on ALL production Glean Chat queries, yielding live user intent insights/trends and task-specific evalsets, while optimizing LLM cost-quality tradeoff and internal UX.

Facebook Video Ads Quality Team (01/2024 – 01/2025):

• Analyzed user behavior/interaction data on FB Reels to design new quality bids, i.e personalized ad scoring/ranking terms incorporating user behavior ML predictions, to optimize user-engagement/revenue tradeoff.

• Designed and built end-to-end new backend logic (Python, C++), data pipelines (SQL, Python), and production ML models (PyTorch + internal tools) for new quality bids, and experimentally verified engagement/revenue impacts & reliability to launch.

• Developed LLM-based (Llama3) pipelines to augment ad classification data for use in quality bids.

Instagram Reels Recommendation Retrieval Team (08/2022 – 01/2024):

• Overall goal: drive user sessions & watch time, by improving candidate retrieval stage of IG Reels recommender system/algorithm.

• Identified, implemented, and integrated cutting-edge ML methods from recent papers (with PyTorch + internal tools).

• Designed & implemented new approaches for rule-based candidate-reel scoring/filtering, in backend (Python/Django) codebase.

Engaged in an immersive three month program to meet and learn from entrepreneurs, executives, and investors at Silicon Valley startups.

• Innovated new product feature (NLP-driven data pipeline generation), and applied product management principles (competitive positioning, adoption barriers, etc) to develop specs/plan, working directly with the CEO of Ikigai.

• Towards this, built (Python) new beginner-friendly search function for data operations in our platform, with NLP methods (semantic similarity, lemmatization, etc).

Researched techniques to improve 3D Object Detection for autonomous vehicles by fusing LIDAR and RADAR data, in Deep Convolutional Neural Network (CNN) methods, in the Kitani Lab at CMU.