Machine Learning Engineer

Megha
Joshi.

Building the future of video generation at YouTube & Google DeepMind. Core contributor to Gemini Omni — Google's next-generation video model.

View my work Schedule a chat Connect on LinkedIn
Currently at YouTube × DeepMind
Focus Video Generation & AI
Education M.S. AI, Penn · B.S. CS & Econ, Yale
Megha Joshi
Education & Experience
Yale University University of Pennsylvania Google YouTube Google DeepMind
Education

Academic Background

Yale University

Yale University

B.S. Computer Science & Economics
Minor in Statistics & Data Science

New Haven, CT
University of Pennsylvania

University of Pennsylvania

M.S. Artificial Intelligence

Philadelphia, PA
Career

Work Experience

Sept 2025
Present

Current

Machine Learning Engineer

YouTube × Google DeepMind — Gemini Omni Video Generation

  • Gemini Omni Core Contributor: Architected the "hillclimb" evaluation suite to benchmark Video-to-Video (V2V) capabilities, identifying critical pre-training gaps that directly informed post-training strategies.
  • EvalSquared: Built a diagnostic tool — now widely adopted across Google DeepMind and YouTube — that reduced eval creation time from days to hours, enabling analysis and correction of large-scale training datasets.
  • Engineered high-throughput Flume pipelines for AI-driven annotations and automated data-matching, streamlining image-video-prompt pairing generation at scale.
  • Built high-quality stylization and reference datasets in Post-Training & DPO while optimizing hyperparameter tuning to refine DPO recipes.
  • Automated the evaluation creation process by mining diverse data sources and building auto-rater pipelines to quantify performance on challenging model capabilities.
Introducing Gemini Omni

Google Blog · May 19, 2026

Introducing Gemini Omni

Create anything from any input — starting with video. Edit naturally through conversation.

Aug 2023
Aug 2025

Google

Software Engineer

Google — Geo Consumer, Mountain View CA

  • Gemini for Maps: Led and productionized a novel video-based Gemini analysis for visual issue detection in Google Maps — from ambiguous problem definition to full-scale implementation — achieving a 10.5× increase in regression detection for 25M daily users.
  • Led applied research in fine-tuning (LoRA) and adapting LLMs for complex visual issue detection, developing innovative prompt engineering strategies including task decomposition and specification grounding.
  • Designed an AI-based bug deduplication system that reduced report volume by ~80% and enabled a 700% increase in operational scale.
  • Co-developed the 2025 all-team AI strategy and presented at the 2024 Geo EngProd Bangalore Summit.
  • Contributed to iOS MapsSDK migration to GMRA, cutting test flake rates from 14–50% to <1% and recovering over 90 SWE hours per month.
Recognition

Awards & Publications

🏆

Google Award

Sweety Silver Award

Google-wide recognition for significantly advancing UI regression detection across mobile applications through the deployment of Gemini-based vision models.

🏆

Google Award

Geo Tech Impact Award

Awarded for the design and implementation of a novel video-based analysis framework using Gemini to automate visual issue detection in Google Maps infrastructure.

📄

Defensive Publication · Lead Contributor

A Method for Identifying User Interface Anomalies in Map-Based Applications Using Prompt Engineering

Novel approach to UI anomaly detection using LLM prompt engineering strategies for large-scale mobile applications.

Read on TDCommons ↗
📄

Defensive Publication · Lead Contributor

Video Analysis for Software Regression Detection Using Generative Models

Framework for applying generative video models to automate software regression detection at scale.

Read on TDCommons ↗