Julian Diaz

Google Search Quality | AI Mode | Model response quality for Local vertical

• Generated vertical-specific training data to improve fanout and model response quality for Local and Directions queries. Achieved more efficient tool calls, improved model responses, and reduced inference steps
• Defined a strategy to improve model performance on complex Local queries via optimized tool calls, thinking, and multi-step reasoning. Built the data generation pipeline and autoraters to produce and validate this data
• Performed a deep-dive quality analysis on model fanouts and responses for Local and Directions, identifying and categorizing key loss patterns, then generated training data to improve model responses
• Experimented with dynamic prompting strategy and implemented dynamic prompts for Local fanout quality improvements

Google Search Quality | LLM Modeling for MapsGuide Topics Model

• Owned prompt engineering efforts for SFT data generation to generate a diverse, high quality final training dataset
• Designed and implemented 10+ autoraters to measure fanout model quality, enabling rapid, data driven iteration on model improvement. Worked with product managers to define requirements and metrics
• Experimented into optimal training data generation methodologies, evaluating trade-offs between fine-tuned vs prompted teachers, LoRA adapter, and Best-of-N selection to increase data quality and efficiency
• Launched 2 model upgrades to the latest Gemini models

Google Search Quality| AI Overviews for Local Vertical

• Defined and Implemented model input ranking algorithms for Local vertical’s key user journeys. The result was consistent inputs for the model to summarize which resulted in consistent responses in AI Overviews
• Ran quality analysis for LLM responses to Local queries. Identified loss patterns for key user journeys, worked with product managers to define optimal local responses, and delivered training data sets for weekly model training
• Enabled inputs from several first party data sources (hotels, restaurant specific data, user reviews, etc,.) which improved factuality and contextuality in model responses

Google Search | The Unified Intent Generation platform

• Supporting next-gen TensorFlow models to run in the critical path for all Search and Assistant queries
• Consolidating intent generation code onto a common set of APIs
• Using those APIs in a unified tooling suite (eval, regression tests, CLI/web debuggers)
• Working with vertical and quality teams to make sure the serving and tooling paths solve the problems they face
• Improving release processes for NLU models and related components

Working on a Demand Side Platform for a Real-Time Bidding Application

• Built an application that will programmatically participate in auctions via ad exchanges and purchase ad space on behalf of Pelmorex's clients
• Built data pipelines for conversion stats, enabling publishers to view website visits and campaign actions
• Expanded user audiences via cookie syncing with Google, Meta, and TTD, improving conversion stats by ~30%
• Developed infrastructure for direct ad space allocation, bypassing the bidding process for predefined contracts

• Developed data analysis and visualization tools allowing plant managers to view production cycle times and identify components that were slowing down production
• Analyzed data and identified components with high cycle variation and recommended setting adjustments which resulted in a 20 % decrease in variation
• Built ETL pipelines for preprocessing IOT sensor data and generating datasets for analysis