Breeana Bellflower
Breeana Bellflower
47.61° N · 122.33° W

Breeana Bellflower · Creator, Architect, Pioneer, Thinker Seattle, Washington

Principal Engineer · Systems & AI

Breeana Bellflower.

I build the platforms other engineers stand on: agent systems, developer tooling, and observability.

Now
Principal Engineer at OneRail
Focus
AI systems · Platform engineering · Developer experience
Also
Mentored for UW's Data Science & Machine Learning certificate programs
Portrait of Breeana Bellflower
№ 01

Selected systems

01 OneRail · AI Systems · Platform

AI Agent Platform

A platform that spawns specialist AI agents to take on real engineering work, reachable through webhooks, Slack, and GitHub. Each agent runs in its own sandboxed container, deployed from a Kubernetes manifest, with zero standing permissions: a virtual key grants scoped MCP access to the data its job requires, and the key dies with the job.

Every new agent gets its own scoped key, so it can only ever reach what its job needs. Each one is created on demand for the single job in front of it, and nothing more.

Agents · MCP · Kubernetes · Slack · GitHub

slackgithubwebhookplatformk8s manifestagentsandboxedvirtual keyMCPsscoped accessspawned per job · zero standing permissions
Fig. 01 · Spawned per job, scoped by a virtual key.

02 OneRail · Developer Experience · CI/CD

Feature Preview Environments

Open a pull request on a feature/ branch and a live container of that branch appears online, wired into staging. Open PRs in other repos under the same branch name, and the containers discover each other and interconnect automatically. Merge, and it all tears down on its own.

In a system with dozens of repos, that meant demoing a feature before it ever touched main, let alone production.

Containers · GitOps · Multi-repo · Staging

repo-a · PRfeature/xrepo-b · PRfeature/xcontainerpreview-acontainerpreview-bauto-linkstagingmerge → containers close out
Fig. 02 · Branch-named containers find each other across repos.

03 OneRail · Observability · OpenTelemetry

Observability Platform

OpenTelemetry APM built around a central collector: every service ships its traces, logs, and metrics to one place, in the same shape, so nothing downstream has to guess who sent what. Tail sampling keeps the traces that matter and drops the noise, so latency, errors, and throughput read the same way across the whole system.

Cheap enough to leave on everywhere, and already the reason behind dozens of bug fixes and performance wins.

OpenTelemetry · Traces · Logs · Metrics · Tail sampling · Collector

svc-asvc-bsvc-ccentralcollectortailAPMevery service exports · keep the traces that matter
Fig. 03 · One collector; keep only the traces that matter.

04 Microsoft · Mobile · Accessibility

Dynamics 365 Field Service Calendar

Brought Outlook-grade calendar UX to Dynamics 365 Field Service on mobile and desktop: three calendar views and booking pages working across browsers, Android, and iOS, with full screen-reader support. Offline mode was the hard part: field technicians lose signal constantly, so the calendar had to keep working alone and reconcile when the connection came back. I also built the CI/CD pipelines behind it and worked from Figma alongside the design org to hold the accessibility bar, all inside a codebase serving enterprise customers at Microsoft scale.

React · Offline-first · Accessibility · CI/CD

Field Service booking calendar on a mobile device
Fig. 04 · Three views, every platform, fully screen-readable.

05 Ace Hardware · Platform · Marketplace

Ace Senior Services

Led the creation of the gig-economy platform that became Ace Home Services: customer, servicer, and admin surfaces, built by three full-stack teams under my lead. A large share of the deliverable work was designed or built by me personally: two of the three surfaces end to end, Stripe payments, Azure auth, accessibility testing, responsive UX. Along the way I introduced Azure Boards and an agile workflow to the company.

Stripe · Azure · Accessibility · Agile

Ace Senior Services web application screens
Fig. 05 · Three surfaces, one marketplace.
№ 02

The record

  1. 01

    OneRail

    Principal Engineer · Now

    Force multiplication as a job description: the AI Agent Platform, the observability stack, the AI gateway, and the C# observability library and monorepo tooling new services build on. I work with peers across the org on the architecture decisions that come before major scale, including readying the platform for multi-cloud. I also designed and led a platform-wide UI overhaul: a container-presenter split enforced by linting, paired with an atomic design system, with a large share of it implemented by me.

  2. 02

    OneRail

    Director of Engineering

    Defined the department's roles and SDLC and bridged product and engineering. Slashed Azure infrastructure spend and cut database CPU to a fraction of its former load while volume doubled. Helped land Lowe's, PepsiCo, and West Marine. Those wins came from delivering on promises under timelines most teams would have missed.

  3. 03

    OneRail

    Senior Software Engineer · Tech Lead

    Built and mentored what was, at the time, the org's highest-performing team. I stayed deep in the code myself, adding 3,800 GitHub contributions along the way.

  4. 04

    Ace Hardware

    Lead Software Engineer

    Led three full-stack teams creating Ace Senior Services, set the engineering bar, and negotiated scope directly with the business. Much of what shipped was designed or built by me personally.

  5. 05

    Microsoft

    Software Engineer

    Rebuilt the D365 Field Service mobile experience: dynamic calendar views, offline mode, custom CI/CD pipelines, and Figma-driven accessibility work with the design org, all inside an enterprise codebase with an enterprise review culture.

  6. 06

    University of Washington

    Instructional Assistant

    Alongside industry work: an instructional assistant for the Data Science and Machine Learning certificate programs I graduated from.

№ 03

Nights & weekends

Nights-and-weekends engineering since 2010. These projects are a small sample of a much longer list.

L·01

System-1 Thinking

AI research · 2026
Shared contextone stateYes / noanswer probabilitiesChoiceanswer probabilitiesScoreanswer probabilitiesshared state · independent questions

A proof of concept inspired by Jev: I trained an open-weight language model with LoRA to return yes-or-no judgments, choices, and scores from answer-token probabilities. Independent questions share the same context, without generating explanatory text. Built from public documentation and research, with a repeatable training pipeline.

Read the write-up

L·02

Inside a Transformer

Interactive · 2026
Tokens→ vectorsAttentionQ · K · VFeed-forwardrefine vectorsNext tokenprobabilitiesan animated walkthrough, step by step

An animated walkthrough I built over a weekend to make transformers easier to follow. Step through embeddings, attention, feed-forward layers, next-token prediction, and the KV cache, with a tiny illustrative model that makes the numbers visible.

Explore the walkthrough →

L·03

Machine Learning

Since 2017
σ(Wx + b) · by hand, since 2017

OCR systems, classifiers, and a pile of side experiments, most of it built on my own time, and a good chunk of it ended up running in production. University of Washington certificates in Data Science and Machine Learning made it official.

L·04

Refinement

Civilization simulator
Refinement civilization simulation running in the browser

A computer-vs-computer civilization simulation, basically an ant farm to watch with friends. Runs are simulated ahead of time, compressed, and streamed in chunks over websockets; every viewer's frame is synchronized to the clock, so the whole room watches the same moment together. Neo-retro shell, chatroom included.

L·05

Custom Operating System

Low-level
Terminal session in a custom operating system

An operating system mirroring Linux's capabilities: terminal commands, process scheduling, memory allocation, a file system, hardware interfacing, plus semaphore locks and user privileges. Debugging at this level is brutal, and no project taught me more about how software actually works.

L·06

Verilog Neural Network

Hardware · Senior design
Hardware neural network diagrams and equations

A neural network implemented in hardware description language: a sigmoid approximation drawn from academic papers for the activation function, and high-precision floating point written from scratch. Simulated neurons matching an equivalent software model. My greatest hours-per-insight challenge to date.

№ 05

Field notes

Astronomy and photography hobbyist, outside whenever the weather allows. Last year I finished a long-running quest: walking the 1,440 miles Frodo covered from the Shire to Mount Doom, logged one mile at a time, most of them on Pacific Northwest trails and shorelines.

Mount Rainier shrugging off clouds above a treeline and lake
A rushing creek below snow-capped peaks
A rock jetty reaching into the Pacific, gulls overhead
A windswept shore pine above the dunes
Storm surf breaking over dark rocks
Sunset over a mountain lake
Tatoosh Range peaks over evergreen forest
Gulls and a wandering dog on wet sand
A camp kettle over an open fire
Beachcombers in the sea haze
№ 06

About the author

I'm a Principal Engineer, and former Director of Engineering, at OneRail in Seattle. My work lives where systems architecture meets AI: agent platforms, model gateways, observability, and the developer experience that holds it all together.

I've been building my own ML models since 2017, mostly OCR and classifiers, on my own time, and a fair amount of it made its way into production. I earned the University of Washington's Data Science and Machine Learning certificates, and later mentored for both programs.

Before engineering I was a classroom educator, and it stuck: explaining things well is half my job, whether the audience is a new hire, an executive, or a room of students. Away from the keyboard I spend time with my wife and child, introducing our little one to the wonder and complexity of the world. I'm also under a night sky, behind a camera, on a long trail, or in a video game.

Languages
TypeScript · Python · React · C#
Expertise
Systems architecture · AI systems · Database management · Data engineering
Databases
MySQL · Postgres · CosmosDB · Elasticsearch · ClickHouse
Cloud
Azure · Heroku · GCP
Human · teaching most of all, and I have a knack for it
Teaching · Leadership · Mentorship · Clear communication

Study

CSU Sacramento · B.S. Computer Engineering

AI-focused senior design · Tau Beta Pi · Hack State competitive programming

University of Washington · Certificates in Data Science & Machine Learning

Instructional assistant for both programs after graduating

Readings that have influenced me

  • Signals and Systems · Oppenheim et al.
  • Operating System Concepts, 9th Edition · Silberschatz, Galvin & Gagne
  • Logic and Computer Design Fundamentals · Mano et al.
  • The Zen of Python · Tim Peters
№ 07

Correspondence

Say hello.

You can reach me any time at

Please don't reach out about sales opportunities. I know it's a challenging job, and I don't want to waste your time!