Senior AI Engineer & Data Scientist

I build chat and agentic RAG systems that people rely on.

For the last three years I have led the AI engineering behind a conversational analytics assistant at Amazon, used by tens of thousands of employees every week. Before that, nearly a decade of production machine learning: recommenders, risk models, and the experiments that proved they worked.

See the work

Currently at Amazon in Seattle, WA.

Portfolio agent

Scripted. Runs in your browser, no model calls.Scripted, no model calls.

LLM systems

Chat and agentic RAG that answer correctly, cite their sources, and stay affordable at scale.

Agentic RAG and multi-agent orchestration
Retrieval and evaluation
Claude, Bedrock, OpenSearch

Production ML

Models that run every day and move a business metric.

Recommenders and risk models
Experiments and causal measurement
Spark, LightGBM, PyTorch, SageMaker

Shipping

Systems owned end to end, and teams helped to ship theirs.

AWS infrastructure with CDK
CI/CD and monitoring
Mentoring and writing

Selected work

Systems I designed and shipped at Amazon. Details are limited to what is public, so the emphasis is on the shape of each problem and what changed once it was solved.

GenAI, 2023 to present

Conversational analytics assistant

An agentic RAG assistant that answers questions about workforce data for Amazon employees, in plain language, with sources.

Lead engineer and architect for the AI system.

Helpfulness, as scored by the judge, rose 20 percentage points.

Anthropic Claude
Bedrock
OpenSearch
Python
TypeScript
AWS CDK
Lambda
ECS

What I built

Retrieval and generation on Anthropic Claude with OpenSearch, serving tens of thousands of people a week, then a multi-agent refactor measured with controlled experiments on cost, quality, and latency.
A custom embedding model and reranker for semantic retrieval, which won 65 percent of head-to-head comparisons against the baseline.
An LLM-as-a-judge evaluation method, plus benchmarking infrastructure for on-demand evals, continuous monitoring, and AI checks in CI.

Machine learning, 2020 to 2022

Warehouse safety risk prediction

A large-scale Spark ML model that flags high-risk conditions in fulfillment centers early enough to act on them.

Built the model and the production pipeline.

Feature engineering over operational and workforce data at fulfillment-network scale.
A scoring pipeline that runs on Spark and feeds intervention workflows.

High-risk events fell 8 percent, a multi-million dollar improvement each year.

Spark
Python
SageMaker
Glue
S3

Machine learning, 2020 to 2023

Personalized report recommender

A LightGBM recommender that puts the right HR report in front of each operations leader.

Owned the system end to end, from data ingestion to serving and monitoring.

Training and serving pipelines for a ranking model over report interactions.
Operational standards for monitoring, security compliance, and reliability that the team adopted.

Serves thousands of operations leaders across the business.

LightGBM
Python
AWS
CI/CD

Machine learning, 2016 to 2019

Attrition risk and manager engagement

Individual-level attrition risk for a workforce of more than a million people, delivered through a website managers actually used.

Built the first model and prototype, then led the production model and its integration.

A predictive model and the feature repository behind it, organized from time-and-attendance and employee data.
A manager engagement website with targeted weekly campaigns.
A/B tests that measured the causal effect of each campaign change.

Targeted campaigns increased employment time by 10 percent across the fulfillment center population.

Python
Scikit-learn
A/B testing
Data warehousing
AWS

Projects

Work outside the day job. The public pieces come first; the rest is in development below.

Open source, 2026Released

Second Opinion - Finance

A second opinion on every money decision, inside Claude Code.

A Claude Code plugin for your own finances. It reads the brokerage accounts you connect and the statements you give it, and works through portfolio, tax, debt, home, and retirement questions together, with company research, valuation, and a whole-market stock screener beside them. The calculations are done by bundled Python scripts on brokerage, SEC, and market data rather than by the model, records stay in local files, and every order needs an explicit approval.

Claude Code
Python
Skills
SnapTrade
SEC EDGAR

View on GitHub

Talk, 2022Released

From notebook to deployed package

A working example of taking an ML notebook to production.

A complete walkthrough from a Jupyter notebook to a tested, versioned Python package with a CI/CD pipeline. Presented at The Data Science Conference in 2022.

Python
CI/CD
Testing
Packaging

View on GitHub

This site, 2026Released

The portfolio agent above

A scripted assistant, no model calls, fully static.

The transcript at the top of this page is a small deterministic agent: weighted keyword and n-gram scoring over a fixed answer set, indexed on the page content itself, with a simulated tool trace and answers that link into the site. The site itself is a Next.js static export on S3 and CloudFront, deployed with AWS CDK.

Next.js
TypeScript
AWS CDK
CloudFront

Next up

One idea runs through all of these. A skill package is the brain of an LLM product, and most of the people it is for will never open a terminal. So each package ships twice: as Claude Code skills, and inside a custom-fitted harness, a native app built around the same skills.

Product, 2025 to presentIn progress
Thesis
A finance assistant that talks to your accounts.
A finance-focused AI assistant for iOS, macOS, and the web, built on an agent platform I designed: domain agents on Strands Agents, a FastAPI service streaming tokens over SSE, tools for market data and brokerage accounts, and sessions in DynamoDB. Runs on Fargate behind Cognito, defined in CDK.
- Strands Agents
- FastAPI
- SwiftUI
- React
- DynamoDB
- AWS CDK
Open source, 2026In progress
Harness Production
Generate a native app around a Claude Code skill package.
A Claude Code plugin that audits a skill package, plans whether to wrap or port it, and generates a standalone harness: a Strands Agents backend streaming over SSE, a short custom system prompt, and a client whose views understand the package's structured outputs. The generated app runs without Claude Code, so a domain package can reach people who will never open a terminal.
- Claude Code
- Strands Agents
- FastAPI
- React
- AWS CDK
Open source, 2026In progress
Music Production skills for Claude Code
An audio engineer, sound designer, and theory advisor in the terminal.
Ten skills that work on a producer's own files: key and chord detection, mix and loudness analysis against a genre reference, click and dropout repair, tempo and groove analysis, arrangement mapping, chord and bassline ideas written to MIDI, and patch recipes for the synths on the user's studio profile. Everything is measured locally by a Python toolkit and audio never leaves the machine. It is the first candidate for a Harness Production build: a native app for producers.
- Claude Code
- Python
- librosa
- MIDI
- Logic Pro

Experience

Nearly a decade at Amazon, told in three chapters that follow the technology rather than the org chart. Before that, data and product work in Europe and the Pacific Northwest. The full history is in the resume.

Amazon, PeopleInsight

Senior Data Scientist

2016 to present, Seattle

Conversational AI
2023 to present
Lead AI engineer for the agentic RAG assistant over workforce analytics, owning retrieval and generation, the evaluation program, and the mentoring of the scientists who ship on it.
- Anthropic Claude
- Bedrock
- OpenSearch
- Python
- TypeScript
- AWS CDK
ML systems and research
2019 to 2023
Owned production ML end to end, including the report recommender and the warehouse safety model, and led research on LLM fine-tuning and Transformer recommenders.
- LightGBM
- Spark
- PyTorch
- SageMaker
- Glue
- Python
Predictive modeling and experimentation
2016 to 2019
Built the first attrition risk model and the data foundation under it, then led the production model, its manager engagement site, and the A/B tests that measured it.
- Python
- Scikit-learn
- SQL
- A/B testing
- AWS

Before Amazon

Business intelligence, data warehousing, and product analytics.

TravelBird
Senior BI Engineer
Amsterdam, 2015 to 2016
Built a product recommendation engine from implicit customer feedback for the website and email campaigns, and ran the A/B test evaluations behind product changes.
Westwing Home & Living
Senior BI and Data Warehouse Developer
Munich, 2014 to 2015
Led the redesign of the central data warehouse and delivered statistical analyses for company leadership.
T-Mobile
Senior Product Engineer and Business Analyst
Bellevue
Prototyped data products from mobile device data and led analytics for customer care operations. Co-inventor on two patents for application diagnostics and quality-of-service comparison.

Education

Master of Information and Data Science
UC Berkeley
Hal R. Varian MIDS Capstone Award
BA, Business Administration
University of Washington
Information Systems and Finance

Patents

Intelligent application diagnostics
US 10,097,434
On-device monitoring that samples per-process performance, keeps the top resource consumers, and reports the filtered metrics for analysis.
Personalized quality-of-service comparison of wireless services
US 9,451,047
Compares carriers for a specific user by evaluating service data from their own device, including services they do not subscribe to.

Working on something with LLMs? Let’s talk.

Open to senior AI engineering and applied ML roles, consulting on LLM products, and collaborations on open source.

brandon.shurick@gmail.com