Michael Brown

I’m an MS/MBA student at Harvard studying how artificial intelligence is trained (SFT, RLHF, RLAIF, RLVR), machine learning systems (containerization, RAG, Kubernetes) and business fundamentals (finance, strategy, marketing, operations). Previously, I worked at Target, where I built simulations to optimize warehousing operations. This summer I was a Summer Consultant at Boston Consulting Group.

Email me Download résumé LinkedIn (opens in a new tab)

Currently studying: Harvard MS/MBA, Engineering Sciences (opens in a new tab)Applied AI and ML, Class of 2027
Previously: TargetSenior System Layout Designer, supply chain

Selected projects

Simulation at Target

Built simulations to optimize warehousing operations across Target’s network of distribution and fulfillment centers, guiding capacity planning and where to spend capital.

Cut operating cost 20% and avoided $10M in capital spending.

One of those projects was the Houston Receive Center, a 1.2-million-square-foot facility Target designed in virtual reality before building it. Target reports the approach saved over $500,000 in materials, and a sorter redesign, checked in simulation, saved nearly $700,000 in construction.

Context: Target Design Engineering, 2022–2025
Built with: Discrete-event simulation, Emulate3D, Python, SQL

corporate.target.com · April 2026 The Real-World Benefits of Virtual Facility Design Read the article (opens in a new tab)

LLM fine-tuning

A 3-billion-parameter language model, fine-tuned to solve math word problems in the voice of Mario. It runs live on the web.

A controlled test of four models showed the accuracy cost came from fine-tuning itself, not from the persona.

Voice similarity rose 13% over the base model; accuracy fell 6 points on 500 held-out problems, the same drop as a plain fine-tune.

Context: Harvard AI Safety (CS 2881) (opens in a new tab) Team of three, 2026
Built with: PyTorch, LoRA, llama.cpp; Read the case study Try the live demo

Breakaway coach

An AI running coach. It syncs your Strava history, builds a fitness profile and writes a periodized training plan that adapts week to week to how your training actually goes.

Designed and shipped solo, end to end: live at breakaway.run with Strava sign-in and a demo anyone can try.

Claude writes each plan, with OpenAI as a fallback. A FastAPI back end with background workers keeps Strava in sync, and the Next.js front end runs on Vercel.

Context: Solo project, 2026
Built with: Python, FastAPI, Next.js, PostgreSQL, Claude API, Strava API

breakaway.run · Live demo Breakaway — AI-powered training plans that adapt to you Click Try Demo to explore with pre-loaded training data. No account needed. Try the live demo (opens in a new tab)

More projects

J-space ablationAnthropic’s global workspace paper (opens in a new tab) found that writing out chain of thought makes a model’s math far more robust to J-space ablation. A preregistered test on Qwen3-4B, using a cheaper logit-lens approximation, didn’t support it: on GSM8K, chain-of-thought accuracy fell 27.7 points against 8.5 for direct answers.
Running studyDesigned a biomechanics study at the University of Minnesota Human Performance Lab, testing 22 custom prototypes. The findings informed footwear design.
Denitrifying reactorBehrens Lab (opens in a new tab) research on farming practices that reduced fertilizer runoff, improving water quality in nearby lakes and rivers.

Experience

2026
Boston Consulting GroupSummer Consultant

Designed the operating model for a state government healthcare program, aligning four agencies on delivering $200M in funding to 4 million residents. Received a return offer.
2022–2025
TargetSenior System Layout Designer

Supply chain engineer in Design Engineering. Lifted building output 15% across 20+ distribution and fulfillment sites while cutting stockouts 3%. Built Python and SQL experiment tracking that cut project lead time 30%, and grew an engineering mentorship program from 33 to 80 people.
2020–2022
PentairEngineering Leadership Development Program

Rotations in R&D, operations and automation engineering. Automated manufacturing and lab processes for $500K in annual savings. Led an offshore engineering team through a new-product introduction, halving the development timeline.
2019–2020
TLC ElectronicsProcess Engineer

With Boston Scientific, stood up a ventilator production line in two weeks during the COVID-19 surge. Rebuilt quoting and cost estimation for more consistent margins.

Education

2025–2027
Harvard UniversityMS/MBA, Engineering Sciences (opens in a new tab)

A joint degree from Harvard Business School and the Harvard School of Engineering and Applied Sciences, focused on applied AI and ML. Coursework includes AI Safety (CS 2881) (opens in a new tab) and ML Operations (APCOMP 215).
2016–2020
University of MinnesotaBachelor of Mechanical Engineering

Magna cum laude. Evans Scholar (opens in a new tab) on a full-ride scholarship, University Honors Program (opens in a new tab), Wallin Scholar (opens in a new tab) and Dean’s List.

Skills

AI and machine learning

LLM post-training (SFT, RLHF, RLAIF, RLVR)
Machine learning systems (containerization, RAG, Kubernetes)
Experiment design
LoRA fine-tuning

Operations

Process simulation (Emulate3D, Python)
Capacity planning
Supply chain optimization
Process automation
Executive communication

Tools

Codex, Claude Code, Antigravity and Cursor
Python, SQL and PyTorch
Data analysis

About

I’m highly curious, which has led me to do research with the University of Minnesota’s Human Performance Lab and Behrens Lab (opens in a new tab).

I love running, music and Nordic skiing. I’m also into photography, camping and specialty coffee.

Contact

mbrown@g.harvard.edu

LinkedIn (opens in a new tab) GitHub (opens in a new tab) Résumé (PDF)