Diamond EdgeCutting edge. Rock solid.

About

Reed Ellsworth

I'm a principal software engineer with 14 years of Android, and the person who does the work at Diamond Edge.

Before Diamond Edge's current apps, I was the sole Android engineer on Weber's Summit Smart Valve grill, owning its UI from architecture to production, and built the companion app that 134,000 users rate 4.6 stars, with live temperature and cook state streamed over BLE and Wi-Fi through multi-hour cooks. Before that, I led Android for point-of-care kiosks on more than 100,000 devices at Outcome Health, and for Sling TV's live-streaming app with 1M monthly Android users, where a single APK came to serve phones, tablets and TVs.

That's the background I bring to an engagement: apps that talk to hardware, fleets that have to update themselves, and products with enough users that a bad release is expensive.

Skills

Android and Kotlin Multiplatform
Kotlin, Jetpack Compose, Compose Multiplatform, Coroutines, Flow, Room, Koin, Hilt, Navigation 3, Material 3, Gradle, BLE, AOSP
Server
Ktor, WebSockets, kotlinx.serialization, Exposed, Postgres
On-device AI
LiteRT, llama.cpp, GGUF quantisation, model conversion, evaluation on real hardware, constrained decoding
Architecture
Unidirectional data flow, MVVM, repository pattern, dependency injection, modularisation
Release
App Store and Google Play submission, App Review, Play Asset Delivery, Apple-hosted Background Assets, store subscriptions
AI-assisted development
Claude daily, with prompt, context and harness engineering
  1. 2020 – nowDiamond Edge, founder. Rimbalo, Diamond Tabletop, Napkin Scores, AI Harness.
  2. 2021 – 2026Weber / Blackstone. Smart-grill UI and companion app; led the move to Compose, Coroutines and Flow.
  3. 2019 – 2021Outcome Health, Android lead. Kiosk fleet: provisioning, remote debugging, OTA updates.
  4. 2014 – 2019Sling TV, principal engineer. Live TV on Android phones, tablets, Fire TV and Android TV; full accessibility.
  5. 2005 – 2014FamilySearch. Desktop indexing client used by volunteers worldwide; first Android prototypes.
  6. EducationMS Computer Science, University of Southern California. MS and BS Mechanical Engineering, Brigham Young University, with an emphasis on AI in design.

Open source

Kotlin Multiplatform libraries I wrote and maintain, on Maven Central and GitHub.

KmLogging

Logging for Kotlin Multiplatform on Android, iOS, JVM and JS, using each platform's own logger. Adopted by hundreds of companies. When logging is off, a call costs one boolean check: the API is inlined and lazy, and each class's tag is computed once.

com.diamondedge:logging · 136 ★

DiamondCharts

A two-series line chart drawn by DiamondCharts

The first charting library for Jetpack Compose: line, bar, stacked, area, pie, scatter and stock charts, now on Compose Multiplatform.

com.diamondedge:charts

Napkin

Buttons, checkboxes and a text field drawn as if sketched on a napkin

A Compose Multiplatform theme in which every component looks hand-drawn. Outlines come from a seeded sketcher, so they wobble but stay put between frames. It's the look of Napkin Scores.

Compose Multiplatform · Material 3

On-device AI

I choose models that run on the phone by measuring them on real hardware: latency, memory, battery, heat, and whether the output is right.

Shipping in Rimbalo

Speech recognition on LiteRT

NVIDIA's Parakeet-TDT converted to LiteRT, running on the GPU or CPU of Android and iOS phones.

Case study · Engineering notes

  • 20 s of speech decoded in under 0.4 s on the CPU of both reference phones, and in 0.10 s on an iPhone SE 3 GPU.
  • Found and fixed an fp16 overflow inside LayerNorm that made two different phone GPUs return the same wrong answer: output similarity went from −0.03 to 0.9997.
  • A Kotlin TDT decoder proven token-for-token against NeMo in CI, and a Kotlin port of NeMo's log-mel preprocessor.
  • Evaluated three on-device LLMs for writing summaries. The one chosen put a false statement in 22% of them, so the generative model came out.

Open source · KMP

AI Harness

An evaluation rig for on-device language models, on Android, iOS and desktop.

Read more · GitHub

  • Claude, Gemini and local GGUF models (llama.cpp through llamatik) behind one streaming interface.
  • A headless runner scores graded suites while recording time-to-first-token, battery, thermal status and peak memory per case.
  • Schema-constrained JSON decoding for the on-device models.

Open source · Python

llm-eval-ci

Graded YAML suites with a pass-rate gate, so a prompt or model change that regresses quality fails the build.

Read more · GitHub

  • An async runner with bounded concurrency and deterministic scorers.
  • Per-call cost and latency, and refusals reported as their own status.
  • The same run-record format the AI Harness writes from the phone.

Have a date you're working against?

Thirty minutes, technical, no deck. You'll leave with a straight read on scope, fee, and whether I'm the right person for it.