AI engineer at ServicePros

I build, research, and teach AI.

I'm back at ServicePros, this time focused on AI: putting it to work in production, finding out what it can and can't do, and helping people use it well. Before that, over seven years of full-stack and DevOps work in Rails, Go, and Rust.

Selected work Get in touch
Now
AI at ServicePros, since 2025
Before
Poll Everywhere, 2023–2025
On the side
Open models & datasets
Also
Photography

01Now

Three sides of the same work.

  1. Implementation

    Taking AI from prototype to production: wiring models into real workflows, with the testing and monitoring that keep them dependable.

  2. Research

    Testing what current models can actually do on real problems, and separating what holds up from what only demos well.

  3. Education

    Helping people understand how these tools work, where they fail, and how to use them with good judgment.

02Work

Selected work.

Most of my AI work is public on Hugging Face: models and datasets. The names are a little silly. The work is earnest.

Fine-tuning · CPT, SFT & DPO · Open weights

SpoomplesMaxx models

A family of open-weight creative-writing and roleplay models, trained with continued pretraining, SFT, and DPO on Qwen, Gemma, Mistral, Granite, GLM, and Seed-OSS bases and released with GGUF and MLX quantizations. It includes Cardmaker, a full fine-tune of Ministral 3 14B that turns a one-line idea into a complete SillyTavern character card.

14K+

downloads of the most popular release, Thrasher 24B (GGUF)

Bases
Qwen · Gemma · Mistral · Granite · GLM · Seed-OSS
Training
CPT → SFT → DPO
Releases
Safetensors · GGUF · MLX

Datasets · Low-resource languages

Ticuna and Kichwa corpora

Data for two under-resourced languages of the Amazon. The Ticuna corpus combines parallel Ticuna–Spanish–Portuguese text, monolingual text, and a lexicon with the dictionary's tone notation, with the orthography tagged on every row. The Kichwa set consolidates and repairs three existing community uploads of a Napo Kichwa–Spanish dictionary into one documented, license-separated dataset.

47K

Ticuna–Spanish–Portuguese parallel pairs

Ticuna (tca)
48K monolingual · 11.5K lexicon entries · 53K instructions
Kichwa (qvo)
13.9K parallel pairs · 6.4K instructions

Datasets · One idea each

Training-data experiments

Small datasets that each test one idea about how data shapes a model. Olivia-Sys rewrites a chat dataset without the hedging, sycophantic assistant voice. repremover-xl puts supervision at the exact turn where a model starts repeating itself.

Olivia-Sys
3,886 conversations in a direct, less assistant-coded voice
repremover-xl
7,281 conversations cut where the model looped, with a rewritten turn

Rust · WebAssembly · Game engine

Nightrunner Engine

A text-adventure engine library written in Rust. It handles parsing and game state for any front end, including the browser through WebAssembly, and is published on crates.io and npm. The demo on this site runs on it.

Screenshot of a Nightrunner text adventure

Earlier

Hugging Face GitHub

03About

About.

I'm a senior software engineer with over seven years across DevOps and full-stack development, and these days most of it is pointed at AI. At ServicePros I work on bringing it into production, researching where it actually helps, and teaching people to use it well.

Before coming back, I spent two years at Poll Everywhere designing presentation streaming with Playwright, FFmpeg, and Go, modernizing code-signing infrastructure, and building geolocation features. In my first stint at ServicePros I led the modernization of legacy systems and rebuilt the CI/CD pipelines.

Outside work I train open models and build datasets, which I publish on Hugging Face, and I keep working on Nightrunner, a text-adventure engine in Rust. I also photograph people and products; that work lives at aimeribaddouh.com.

Portrait of Aimeri Baddouh

Experience

  1. 2025 – Now

    ServicePros Installation Group

    AI implementation, research & education

    Building Claude workflows for the team, using Amazon Bedrock to bring AI into complex processes like contract verification, and rolling out vendor AI tools that integrate deeply with our chat, inventory, and scheduling systems. Also teaching the team to use all of it well.

  2. 2023 – 2025

    Poll Everywhere

    Led development of presentation streaming solutions, implemented geolocation features, and mentored junior developers on modern Rails patterns and API best practices.

  3. 2019 – 2023

    ServicePros Installation Group

    Overhauled CI/CD pipelines, upgraded legacy code to modern standards, and integrated third-party tools including APM monitoring, Google Maps, and OAuth authentication.

Toolbox

AI platforms
Claude, Amazon Bedrock, Hugging Face
Model training
Transformers, TRL, Unsloth, Axolotl, PEFT & LoRA
Inference
vLLM, llama.cpp (GGUF), MLX
Languages
Python, Ruby, JavaScript, TypeScript, Go, Rust, Elixir
Web
Rails, Hotwire & Stimulus, React, Phoenix, CSS & SCSS
Infrastructure
Docker, AWS, CI/CD, GPU training on RunPod and Vast.ai
Craft
API design, accessibility, UI/UX design

04Writing

Writing.

Notes on software architecture, game engines, and whatever I'm building.

All writing

05Contact

Get in touch.

Interested in working together, or want to talk about AI? Send a note and I'll get back to you promptly.

Hugging Face
huggingface.co/aimeri