Beckett Dillon
Contact
Menu

Background: an interference field of six wave sources. Move the pointer to add a seventh, and click to send a pulse. Ink lines join points of equal phase, the wash shows amplitude, and small rings mark phase vortices where the wave cancels to zero.

Applied AI engineer · open to new roles

Patterns in soil, sound, wood and code

An AI engineer who started in habitat restoration and audio, and builds systems that work with the grain of a problem. Most recently I put generative AI into production at Adobe.

See the work Résumé

One way of working, four mediums

Water finding its way downhill, a mix settling in a room, grain running through a board, a model settling into a loss basin. I learn how a medium moves, then build with it instead of against it.

  1. Soil and water

    Riparian and wetland restoration with the Washington Conservation Corps, 250,000+ trees planted; habitat, irrigation and crews at Apricot Lane Farms.

  2. Sound

    More than 300 commissioned songs, live event audio, and the speech and music pipelines behind a 24-hour channel.

  3. Wood and print

    Heirloom furniture built from blueprints, jigs and glue-ups; parametric parts printed in PETG and tested by the coupon.

  4. Models and code

    Human-feedback training on frontier LLMs, 46 open models, and generative AI in production at Adobe.

Selected work

Production work inside large organizations, alongside the systems I build to learn a new medium. Every number here comes from a delivered project or a public count.

  1. 2025–26AdobeBaptist Health · Logitech · Delaware North · Fanatics

    Shipping AI inside four enterprise stacks

    At Adobe I embed with a customer's engineering and creative teams and integrate generative models into the systems they already run, then carry what we learn back into Adobe’s own products and tools. Most of the work is integration engineering: identity, asynchronous APIs, metadata contracts and the permissions real organizations have.

    enterprise customers
    4
    enterprise customers
    Creative Cloud plugins: Illustrator, Photoshop, Premiere Pro in pilot
    3
    Creative Cloud plugins: Illustrator, Photoshop, Premiere Pro in pilot
    rescope found at Baptist Health, 25 items to about 250
    10×
    rescope found at Baptist Health, 25 items to about 250

  2. 2023–26IndependentANIMA, a live biomimicry agent and workbench

    Biomimicry that cites its work

    ANIMA takes an engineering problem, finds organisms that solve the same function, checks whether the strategy transfers to your scale and sizes it with real physics. An autonomous agent plans the research, runs tools in parallel, delegates to subagents and audits its own citations before it answers.

    curated biological strategies
    239
    curated biological strategies
    cited DOIs, all verified against Crossref
    314
    cited DOIs, all verified against Crossref
    physics calculators with cited assumptions
    21
    physics calculators with cited assumptions

  3. 2023–26Independent and Invisible TechnologiesHugging Face · GitHub · frontier labs via Invisible

    Training, merging and porting models

    Before I wired models into enterprise stacks I trained them. A year of human-feedback work on frontier LLMs, 46 open models and 7 datasets of my own, and a port of a new brain-inspired architecture to Apple silicon.

    models released
    46
    models released
    all-time downloads
    188k
    all-time downloads
    training throughput after the MLX port
    ~2×
    training throughput after the MLX port

  4. 2026IndependentSlopSoup TV on YouTube

    A television network that runs itself

    SlopSoup TV is a 24-hour pixel-art channel run end to end by software: writing, casting, voices, music, scheduling and the live broadcast. It is my test bed for long-running AI systems that have to stay up, stay on budget and stay in tune.

    on air, on a fixed published grid
    24/7
    on air, on a fixed published grid
    shows with their own bibles and critics
    20
    shows with their own bibles and critics
    test files, including a simulated 24-hour playout
    111
    test files, including a simulated 24-hour playout

  5. 2024–25Doner and YamamotoACE Hardware · Stellantis · Tylenol · Rush Street Interactive

    Turning brand and legal rules into code

    At two Stagwell agencies I took rules that lived in PDFs and people's heads, from a retailer's voice to automotive legal disclaimers, and turned them into pipelines that apply them every time.

    less copywriting time for ACE Hardware
    60%
    less copywriting time for ACE Hardware
    faster campaign delivery
    40%
    faster campaign delivery
    less manual input on digital signage
    70%
    less manual input on digital signage

  6. 2024–26IndependentGitHub

    Open tools for running AI locally

    GraphRAG-Local-UI made Microsoft’s GraphRAG usable with local models, with an indexing, prompt-tuning and query interface. It is one of several open tools I maintain for running retrieval and agents on your own hardware.

    stars on GraphRAG-Local-UI
    2,321
    stars on GraphRAG-Local-UI
    forks
    291
    forks
    stars on Vodalus
    195
    stars on Vodalus

Soil, studio, systems

Read it like a core sample, newest on top. I spent years with my hands in the ground and in workshops before I wrote production code, and the habits came with me: measure twice, respect the system you are working in, leave it better.

  1. Systems · Production AI integrations and agentic automation inside large organizations

    1. Senior Forward Deployed AI Engineer and AI Consultant · Adobe

      Nov 2025 – now · Remote

      Embedded with enterprise customers and with Adobe’s own product teams: integrating Firefly Services, AEM and GenStudio into customers’ systems, and turning what the field taught us into internal tools, demos and product changes.

    2. AI Solutions Engineer and Automation Developer · Doner

      Feb 2025 – Sep 2025 · Contract · Detroit, Michigan

      Turned brand, legal and production rules into pipelines for ACE Hardware, Stellantis, Tylenol and Rush Street Interactive campaigns.

    3. AI Engineer · Yamamoto

      Aug 2024 – Sep 2025 · Minneapolis, Minnesota

      Brought generative AI, AR and spatial computing into campaigns with the innovation team, alongside the Doner role.

  2. Models · Training frontier LLMs with human feedback; an independent studio for open models, tools and craft

    1. Founder, independent studio · Interwoven Arkitech

      Jan 2022 – now · Remote

      My independent practice for design, open-source AI, software and craft: tools I build to learn a new medium, and that other people use.

    2. Advanced AI Quality Analyst and Data Trainer · Invisible Technologies

      Sep 2023 – Aug 2024 · Remote

      Trained and evaluated frontier language models with human feedback.

  3. Studio · Audio production degree, 300+ commissioned songs, live event AV

    1. Audio Visual Specialist · Pinnacle Live

      Jan 2022 – Oct 2023 · Asheville, North Carolina

      Live audio, video and lighting for conferences, classes and hybrid events; rigged stages and coached presenters.

    2. Verified Artist · Songfinch

      Jan 2019 – Sep 2023 · Remote

      Wrote, recorded, mixed and mastered 300+ commissioned songs from customers’ stories, each in four days.

  4. Workshop · Heirloom furniture built from blueprints, jigs and glue-ups, quality inspection

    1. Production and Assembly · Geek Chic Custom Furniture

      May 2015 – Dec 2018 · Seattle, Washington

      Built heirloom gaming tables from blueprints: jigs, glue-ups, machining and final inspection.

  5. Soil · Habitat restoration crews, 250,000+ trees planted, a biodynamic farm, irrigation, training volunteers

    1. Landscape, Habitat Restoration and Construction · Apricot Lane Farms

      Jul 2012 – Mar 2015 · Moorpark, California

      Led habitat restoration on a regenerative farm, including a riparian pond and forest; ran irrigation and trained volunteers.

    2. Assistant Supervisor · Washington Conservation Corps

      Oct 2010 – Jan 2012 · Vancouver, Washington

      Helped lead an AmeriCorps crew in wetland restoration, trails, and flood and wildfire response. Over my restoration years I planted 250,000+ trees, up to 20,000 a day.

Full résumé with every role

Model work

I know how models behave because I have trained them: a year of human-feedback work on frontier LLMs, then fine-tunes, merges and quantized builds of my own, and a port of a new architecture to Apple silicon.

Training with human feedback

Invisible Technologies · 2023–24

Reviewed frontier model outputs, named the failure patterns and wrote the feedback that steers them. Prepared training data and wrote evaluation rubrics.

Open releases

Hugging Face · 2023–now

Fine-tunes on Mistral and Hermes, mixture-of-experts merges, early Jamba hybrids and the datasets behind them, released as LoRA, GGUF, MLX, GPTQ and AWQ.

New architectures

BDH-MLX · 2025–26

Ported Pathway’s Dragon Hatchling from PyTorch to MLX for about twice the throughput, then built recursive and mixture-of-experts variants.

Models
46
Datasets
7
Downloads
187,876
On community builds
137,519

Base modelMy fine-tune or mergeMy buildCommunity build

Mistral-7BHermes-2-Pro-Mistral-7BSciPhi-Mistral-7B-32kJamba-v0.1Meta-Llama-3-8BMistral-Small-24BANIMA-Phi-Neptune-Mistral-7BANIMA-Cognitive-Mistral-v1ANIMA-NectarNexus-IKM-Mistral-7B-v5-instructionNexus-IKM-Mistral-Instruct-v0.2-7BNexus-4x7B-IKMNexus-IKM-Mistral-7BMistral-v0.2-Nexus-Internal-Knowledge-Map-7BEinstein-IKM-v1-7BNexus-IKM-Hermes-2-Pro-Mistral-7BNexus-IKM-RolePlay-StoryWriter-Hermes-2-Pro-7BANIMA-SciPhi-7B-32k-v1Jamba-UltraInteract-Instruct-1BJamba-Nexus-4xMoEJamba-900MJamba-v0.1-Claude-ChatLlama-3-IMPACTS-2x8B-64kGlyphstral-24b-v1
36 builds shown from a Hugging Face snapshot taken 2026-11-04. Dot size follows all-time downloads.
Inside the model lab, with a Transformer and BDH side by side

Built outside of work

Smaller things I make to learn, to help people I know, or because the problem would not leave me alone. Each one is labeled with where it stands.

  • No. L12026Science-backed training app, built for myself

    Live

    Functional Bodyweight Workout

    A training app for strength, mobility and capability with nothing but bodyweight and a set of stackable resistance bands. Every choice in the program traces back to peer-reviewed exercise science, and the app counts a partial session as progress, not failure.

  • No. L22025–26Image protection for artists

    Working build

    Ghostprint

    Adds a perturbation people cannot see to an image so that AI models misread it, to protect artists’ work from unauthorized training. Tested against ResNet, ViT, CLIP and the Stable Diffusion VAE within a set perceptual budget.

  • No. L32025Offline wilderness and first-aid guide

    Live

    Flint

    A survival and first-aid companion for places without signal. It offers to download a local AI model to the device on first launch, so the assistant keeps working offline. Built from years of backcountry crew work.

  • No. L42025Civil-rights companion for encounters with authorities

    Live

    Firewatch

    A free app that helps people know their rights, document an encounter safely and reach legal help fast. It works offline and in more than a dozen languages.

  • No. L5202624-hour AI music channel

    In progress

    Slop FM

    The music sibling of SlopSoup TV. Songs written and generated around the clock with ACE-Step, lyrics timed with Whisper and pixel music videos cut to the beat, on CPU at no cost.

  • No. L62025–26Parametric design and multi-color printing

    Working build

    Printed parts

    Functional and decorative parts designed in code and printed in PETG on a Bambu Lab printer, from parametric lattices to a lantern, a guitar body and masks painted for four-colour prints.

I also keep a research notebook on signal analysis and long-term memory for language models. It is exploratory and lives on GitHub.

How I work

The same four steps on a restoration crew, in a recording studio and inside an enterprise integration.

  1. 1

    Learn the medium

    Spend time where the work happens, on a riparian slope, at a mixing desk or in a production team’s queue, until I know how the material behaves.

  2. 2

    Find the pattern

    Name the rule underneath: how water crosses a site, how a brand’s legal line is formatted, where a model’s errors cluster.

  3. 3

    Build the instrument

    Make the smallest tool that works with that pattern on real inputs, then prototype and iterate as fast as the problem allows.

  4. 4

    Hand it over

    Docs, tests and training, so the people who live with it can tune it without me.

Ask about my work

Answers come only from this site, and each one shows the passages it used. Search runs instantly. Load the model and a 2-billion-parameter LLM runs on your own GPU, so nothing you type leaves your machine.

How this runs
  1. 1Your question is matched against every passage on this site with BM25, in the browser.
  2. 2The top five become the model’s only context, numbered so it can cite them.
  3. 3LFM3-Omni-2B, 4-bit GGUF runs over WebGPU with Liquid AI’s runtime, checked against a pinned checksum before it runs.
  4. 4The model can call tools on this page: run the pipeline demo, scroll to a section, open a case study.
  5. 5Without WebGPU, a small server function answers from the same passages, rate-limited, with its key kept server-side.

Search in your browser

about 1.6 GB, downloaded once and cached

Work together

Open to new roles. I am looking for applied AI work in climate, agriculture and conservation tech, in creative and audio tools, and anywhere a hard real-world problem needs someone who learns the domain fast. Email is the quickest way to reach me.

[email protected]LinkedIn
esc

↑↓ to move ↵ to open tab to ask the model