LLM training · Fine-tuning · Inference

When prompting
isn’t enough.
Train it on
your data.

Measured properly. I’m Saqib: I’ve trained models on healthcare text and imagery, and I built Strata to run big Mixture-of-Experts models in a fraction of the memory.

Studio portrait of Saqib Ejaz in a charcoal blazer
Scroll to exploreTraining · Fine-tuning · MoE inference
01 / Where I come in

Past the
prompt ceiling.

Prompts and RAG first. Training when they stop paying.

01 / CEILING+

“Prompting has hit a wall.”

Train on your data when better prompts stop moving the numbers.

02 / SENSITIVE DATA+

“Our data is regulated.”

Trained on healthcare text and imagery within HIPAA and SOC 2.

03 / HARDWARE+

“The model won’t fit our machines.”

Strata pages MoE experts from SSD, so big models run in less memory.

04 / PROOF+

“Better how? Says who?”

Before-and-after evals. Every gain comes with its quality cost.

05 / WORKFLOW+

“A model alone doesn’t do the job.”

Wire it into agents that answer questions and book appointments.

02 / Work behind the words

Trained, shipped,
published.

60 GB

MoE model (GLM-4.5-Air) loading and generating in under 12 GB peak memory with Strata

8+

years shipping production software, end to end

Healthcare AIHIPAA · SOC 2
Model training

Trained models on healthcare text and imagery to power an in-house agentic workflow that answers patient queries and books appointments.

Amperor Strata ↗PyPI · amperor-strata
Creator

Runs Mixture-of-Experts models bigger than your RAM on Apple Silicon by paging experts from SSD. OpenAI-compatible server. Published on PyPI.

Qwen3.8-27B ↗Personal project · MacBook M5
Training

Trained Qwen3.8-27B on a MacBook M5: runtime training to sharpen browser tasks and decision making, then weight training to turn it into a solutions architect.

UltralyticsComputer vision
Full stack engineer

YOLO object detection products, an end-to-end Flutter app with SDK integration, and Kubernetes deployments.

Strata figure: measured by Amperor on an Apple M5 Max (see strata.amperor.ai/methods).
03 / Stack

Data to deployment.

Training is half of it. Running it cheaply is the other half.

Model trainingFine-tuningQwen3.8-27BMLXMixture-of-ExpertsApple SiliconPythonOpenAIAnthropicRAG pipelinesAgentic workflowsAWSAzureDocker
04 / Pick your flavour

Same engineer.
Your problem.

Choose the one that sounds like yours. Each page is the short version for that job.

05 / Let’s talk

Tell me about the model.

What it should do, what data you have, what it runs on. I’ll tell you whether training is worth it.