Case Studies &
Technical Reports

Each document captures a real problem, what was tried first, what failed, and what ultimately worked — drawn from two decades of building ML systems at scale in e-commerce, enterprise platforms, and document AI.

Adobe

Head of AI & Data Science · Document Cloud & Experience Platform

Building Segmentation AI from Zero to Enterprise Launch: How Architecture Redesign, Cost Discipline, and an Intent-Driven Product Model Scaled Audience Discovery at Adobe

Inherited a leaderless team with a delivery deadline and a prototype capped at 2M profiles. Replaced GPU-bound deep learning with distributed Spark, built a Bayesian ranking system so marketers could search by campaign intent instead of browsing 40 clusters, added Approximate Nearest Neighbor expansion, and worked with the pricing team to hit 75% margin — scaling to 5B profiles while cutting cost by over 70%. Deployed with Verizon, Home Depot, and CBSi.

From 20% to 60% Precision in Six Months: How First-Principles Thinking and Data-Centric AI Transformed Document Understanding at Adobe

A document understanding system stuck at 20% precision wasn't a model problem — it was a data problem. First-principles diagnosis, systematic data-centric interventions, and iterative error analysis tripled precision in six months.

Salesforce

Head of AI · Community Cloud & Service Cloud

Driving Engagement in Enterprise Social Networks: How Four ML Systems Transformed DAU/MAU

Given the mandate to improve DAU and MAU, diagnosed four failure modes and built four coordinated systems: feed relevance ranking, multi-entity recommendations, Q&A answer ranking, and spam detection.

Making Enterprise AI Stick: How Solving Two Adoption Failures Transformed Service Cloud AI

Customers couldn't configure chatbots; agents didn't trust answer suggestions. Two different failures requiring two different fixes — vertical defaults for the chatbot, dual-index retrieval for the answer engine.

Multi-Objective Recommender Systems for Enterprise Social Networks

Mathematical formalization of the unified latent space, composite engagement score, and two-stage adaptive target optimization for multi-entity recommendations across heterogeneous organizations.

Walmart

Head of Search Science · 80+ person org

From Diagnosis to Impact: How Systematic Problem Decomposition Drove +23% Revenue Lift in E-Commerce Search

60% of search failures were business problems misclassified as technology problems. A diagnostic framework that separated relevance, ranking, and demand-supply gap issues before building models.

Building Diagnostic Analytics to Focus Applied Science Teams

The measurement and analytics framework behind the diagnostic approach — how to instrument a search system so failures are classified before resources are committed.

Quantifying the Demand–Supply Gap in E-Commerce Using Topic Models

Using topic models to identify query spaces where customer demand exists but catalog supply does not — closing the gap between what customers search for and what the marketplace offers.

Controlled Experiments for Decision-Making in E-Commerce Search

Guidelines for A/B testing in e-commerce search from experiences at WalmartLabs: visit-level, query-level, and item-level bias; OEC selection; non-parametric testing for skewed metrics; holiday seasonality; and financial budget constraints. Published in IEEE BigData 2015.

Elance

Head of Data & Machine Learning · now Upwork

Experimentation Design for Two-Sided Labor Market Matching

A/B testing in a marketplace where treatment on one side affects outcomes on the other. Designing experiments that account for interference, network effects, and equilibrium shifts.

eBay

Search Science Lead

Learning to Rank in E-Commerce Marketplaces: Why Target Design and Feature Engineering Beat Model Complexity

Replacing a 15-year heuristic ranking system with ML. The key was not the model — it was extracting tacit knowledge from the heuristic into features and designing price-normalized training targets. Includes GBT-to-C compilation for production serving.

When Better Is Worse: How a Failed Image Quality Experiment Revealed What eBay's Buyers Actually Value

Boosting high-quality images in search made pages look better — and revenue dropped. In eBay's core C2C categories, "low quality" mobile photos are authenticity signals. Buyers optimize for trust, not aesthetics.

Cross-Company

Walmart · eBay

Price as Structure, Not Feature: Three Problems in Commerce Search Where Getting Price Wrong Breaks Everything Else

Price defines consideration sets, conditions relevance, and determines marketplace value allocation. Three interventions — each producing >2% revenue lift — that form a structural hierarchy every commerce ML system should address.