Projects

Selected anonymized consulting work, plus relevant examples from previous full-time roles.

AI Coding Workflow Scale-Up and Architecture Review

Context

Healthcare startup — platform and shared services team (consulting engagement, 6 months, extended from 3)

Challenge

• The team had useful AI coding workflows, but they required too much manual babysitting
• Pull request volume was increasing, but review capacity and merge flow had to keep up
• Large AI-generated PRs were harder to review and slower to merge
• The team needed a cheaper, more repeatable way to run multi-step AI workflows
• The platform needed an architecture review and a prioritized roadmap before the next phase of work

Solution

• Completed a full architecture review and prioritized roadmap in the first month, covering eight domains including API and UI design, role-based access control, deployments, and data ingest
• Moved the workflow from a manually watched local process into scheduled/background execution
• Helped turn nested prompt-and-skill chains into more deterministic workflows
• Wrote a TypeScript pipeline library that composes LLM calls with shell commands and other logic, choosing the model and effort level for each call
• Added enough structure to measure throughput, merge flow, and cost per feature
• Wrote a harness strategy for standardizing on Claude, Codex, and OpenCode across the company: shared skill libraries, tests that verify skills work in all three, and support for all three in the background coding workflow
• Explored using AI workflows for metrics and operational work in the final weeks

Impact

• In the first four months after launch, PR volume rose 111% over the prior three-month baseline
• The AI coding workflow system authored about 70% of that increase
• Merged lines of code rose 40% over the same period, with the system producing 38% of the gain
• Routing workflow steps to cheaper models reduced cost per feature by 30%
• The architecture roadmap gave the team a prioritized plan across eight domains and a clear starting point for implementation

ML Orchestration Strategy & Roadmap

Context

Mid-sized social gaming company — AI/ML team (consulting engagement, 2-month project)

Challenge

• The team was running Dagster, Airflow, and Argo at the same time
• Engineers had friction moving jobs from ad-hoc development to production
• Dependency management was brittle
• GPU scheduling caused frequent failures
• Observability was inconsistent across batch jobs

Solution

• Worked with ML engineers to review the ML orchestration landscape and develop a 1–2 year technical roadmap with measurable goals for iteration speed, job success rates, GPU utilization, and platform reliability
• Defined a target architecture built on Ray and Dagster, with projects ranging from dependency isolation to automated job testing and Airflow-to-Dagster migrations
• Prioritized initiatives across developer experience, resource efficiency, and system reliability
• Wrote a GPU strategy to address job failures from GPU scarcity and low utilization: reservations, queued provisioning, scheduling, and multi-region failover

Impact

• Delivered a roadmap covering 10+ concrete projects, from dependency isolation to automated model deployments
• Gave leadership clearer technical direction, quarterly milestones, and a better basis for resource allocation
• Reduced uncertainty around platform evolution without forcing a premature rewrite

Website Latency Fixed in Four Hours

Context

Small CPG startup — ecommerce site (one-day consulting engagement)

Challenge

• The client’s website had intermittent latency and availability problems, with no clear lead toward a root cause

Solution

• Traced the failure mode across the request path
• Identified DNS configuration as the source of the problem
• Corrected the configuration and verified the site behaved reliably afterward

Impact

• Fixed the issue in roughly four hours
• Restored reliable website access without a rebuild or migration
• Gave the client a clear explanation of what failed and why

Personalization Runtime Analytics

Context

Twitter — Home ML Systems (previous full-time role)

Challenge

• Twitter served billions of personalized timelines per day
• Engineers, data scientists, and product teams had limited visibility into why specific tweets appeared in production timelines
• Teams could not easily understand how model changes affected real users

Solution

• Built an analytics platform that logged candidate sourcing, ranking decisions, filtering logic, and multiple ML model scores per tweet
• Collected data inside runtime personalization systems, queued it through HDFS, and landed it in GCP BigQuery for internal analysis
• Built a Timeline debugger that annotated individual timelines with model scores, features, and ranking metadata

Impact

• Helped engineers, researchers, data scientists, and product managers understand personalization behavior in production
• Made it easier to investigate “why am I seeing this?” questions and validate model changes against real timelines
• Reduced knowledge silos around how timeline personalization worked

User Sampling Pipeline

Context

Netflix — Personalization Infrastructure (previous full-time role)

Challenge

• Netflix personalization pipelines needed diverse, representative training datasets for global models serving roughly 100 million members
• Existing sampling approaches lacked precise control over country and tenure coverage
• Teams needed a reliable way to over-sample strategically important segments

Solution

• Built a user-sampling service that produced balanced datasets of 50M+ users by country and tenure
• Supported configurable sampling strategies for representative global coverage and targeted over-sampling
• Integrated the service with downstream ML training jobs that needed consistent, high-quality datasets
• Built a page simulation service for offline ranking experiments, filtering model variants before online tests

Impact

• Enabled global personalization models in place of separate per-region models
• Gave machine learning engineers a dynamic way to express training data requirements in code across dozens of models
• Let teams filter down model variants offline before online tests

Making Netflix’s API Multi-Region

Context

Netflix — Edge Platform (previous full-time role)

Challenge

• Netflix was expanding to Europe, which required serving its API from multiple AWS regions
• Device traffic needed to reach the right backend services and move between regions as needed
• The edge needed to stay operable and resilient while carrying heavy device traffic

Solution

• Built API traffic routing infrastructure for the European launch, including global DNS and a new HTTP-based device redirect protocol
• Led the migration of about 50% of western-hemisphere traffic into a new AWS region
• Added the ability to fail traffic over from one region to another through cross-region proxying
• Co-authored Zuul, Netflix’s edge proxy, which routed about 100k requests per second of device traffic to backend services, with a focus on operational hardening, runtime insights, and resiliency features

Impact

• Enabled Netflix’s European launch across hundreds of device types
• Gave Netflix region-to-region failover for its API traffic

Grepzilla: Real-Time Edge Traffic Queries

Context

Netflix — Edge Platform (previous full-time role)

Challenge

• Engineers investigating production issues had limited visibility into live device traffic
• Understanding how devices actually communicated with Netflix services required digging through logs

Solution

• Built a real-time event processing system for about 100k requests per second of Netflix edge traffic
• Made traffic queryable by country, status code, customer ID, and many other dimensions

Impact

• Used by hundreds of engineers for debugging and information gathering
• Became a primary tool for investigating production issues and spreading knowledge about how devices communicate with Netflix services

Personalized Video Re-Architecture

Context

Comcast — Video Personalization (previous full-time role)

Challenge

• Content-personalization ranking logic was embedded inside browse and search services
• The tight coupling made it harder for machine learning engineers to improve models independently
• Search-service constraints slowed personalization work

Solution

• Ran an architectural A/B test to evaluate latency, availability, and error rates across caching and page-size configurations
• Added fallback behavior so video discovery could degrade gracefully when personalization systems were unavailable
• Led the launch of the personalized “For You” video experience, including capacity planning, load testing, and improved metrics
• Built a video metadata caching system that sped server startup by 4x and cut deployment time by 50%
• Built a pipeline to rank popular content and enable dynamic popularity-based sorting anywhere within the video experience

Impact

• Launched the new personalized video experience to about 20 million Comcast video customers
• Increased customer video engagement by roughly 10%
• Sped server startup by 4x and cut deployment time by 50%