What we do
Custom AI, Built End-to-End
Data, training, serving, and the agents on top. One stack, built and operated by the same people, so nothing falls down the gap between two vendors.
LLM Fine-Tuning
A model trained on your data, sized for your budget, and measured against the baseline you are already paying for. SFT, DPO, ORPO and every major preference-optimization method, on hardware from a single GPU to a multi-node cluster.
A 4B specialist we trained beat a 27B generalist at its own task by 3× on recall and ran ten times faster, cutting the serving footprint that goes with it.
Dataset Curation
Your training data decides whether your model works. We build conversation datasets, preference pairs and corpus pipelines to production standard, documented and versioned so an auditor can follow every row back to where it came from.
We audited a 14,816-pair dataset, derived a reproducible quality detector that flags 70% of machine-written text against 0.3% of human, and ship code datasets verified by execution.
AI Integration & Agents
AI that fits how your business already runs. Private local inference for regulated data, hybrid cloud where it pays, and production agents in Slack, Teams, Discord or your own product.
NikuNiku900 serves image generation to Discord and Twitch on our own GPUs, and the same stack runs our training queue and inference fleet around the clock.
Cloud Native DevOps
The infrastructure underneath it all: automated delivery pipelines, reproducible environments and orchestration you own outright, with no per-seat licence and no vendor holding your weights.
Witchgrid, our open-source orchestration layer, runs the production inference fleet across every GPU we own.
Our Portfolio
Software We've Built
Infrastructure & Tooling
Merlina trains the models, Witchgrid serves them, egirl puts agents on top. One open-source stack, and we run our own business on every layer of it.
Merlina
Schneewolf Labs
Witchgrid
Schneewolf Labs
egirl
Schneewolf Labs
flammen.ai
Schneewolf Labs
Products & Games
Consumer products we own and operate. They put the platforms above under real load every day, at our expense rather than a client's.
NikuNiku900
Schneewolf Labs
Yotis Studios
Yotis Studios
Raifu Wars
Yotis Studios
Elements of Estoria
Yotis Studios
About
Who's behind the work
Nick Beerbower runs Schneewolf Labs. Penn State alum, formerly at Anaconda, post-training contractor to Zyphra on tool-calling SFT and evaluation, and a working contributor to open-source AI: on Hugging Face, known for the Gutenberg DPO tunes, models trained to write like a person rather than like a language model.
The research is public. Five Hugging Face accounts hold the personal research line, our own models, flammen.ai's character models, the Hemlock programming language and Yotis Studios. Hundreds of open models, dozens of open datasets and over half a million downloads between them, alongside dozens of repositories on GitHub. Every method we sell is one we published first, so you can read it before you buy it.
The tooling is public too, and it exists because we needed it. Merlina is the interface we run our own fine-tuning through, Witchgrid orchestrates the inference fleet, and egirl is the agent layer on top. Our training queue and inference fleet run on them daily, on hardware we own. They hold up because our own work depends on them, not because they passed a demo.