Machine Learning Engineer for MAGIC AI

Job ID: 40050360

Budget: $10 – $400 USD

Job Title: MAGIC AI – Official Model Trainer

Status: Contract (with opportunity for long-term collaboration)
Compensation:

$200 competition prize

$400 official trainer contract

Public credit on MAGIC GitHub & HuggingFace repositories (MAGIC + MageV1)

About MAGIC AI

MAGIC is a new AI architecture currently under active development. We're looking for a dedicated and highly skilled machine learning engineer to serve as the Official MAGIC Trainer, working directly on the training pipelines, performance tuning, and dataset refinement for the MAGIC and MageV1 foundation models.

This role is ideal for someone who enjoys experimental architectures, training large models from scratch, and pushing performance boundaries through iterative improvements to the training loops.

How Selection Works

To be considered for the role, applicants must participate in and win the current competition:

Competition: https://www.freelancer.com/contest/Training-MAGIC-AI-2663043/

Prize: $200
Goal: Train the best-performing MAGIC Base model
The winner of this contest will automatically be offered the official training position.

Position Summary

As MAGIC’s Official Trainer, you will:

Work directly with the developer of MAGIC on model training strategy

Improve and extend the core training script (training_magic.py)

Design, run, and optimize training loops for MAGIC Base and subsequent versions

Handle preprocessing, batching, scaling strategies, checkpoints, and evaluation

Experiment with hyperparameters, optimizers, scheduling, and architecture variants

Provide input on dataset sourcing, creation, and curation

Collaborate on training reproducibility and documentation

Help shape the official MAGIC training standards for future models

Responsibilities

Maintain and enhance MAGIC’s training pipeline

Run training experiments and document performance results

Implement improvements to stability, speed, and model quality

Optimize for GPU/multi-GPU setups

Collaborate on shaping the evolution of MageV1 and future MAGIC versions

Contribute to best practices and training documentation

Ensure reproducible training via versioning and clean code updates

Requirements

Technical Skills:

Strong experience with PyTorch (preferred) or similar deep learning frameworks

Hands-on experience training transformers or other large neural architectures

Strong understanding of training loops, optimization, scheduling, and GPU efficiency

Experience with large-scale data processing and dataset pipelines

Familiarity with multi-GPU, distributed, or accelerated training frameworks

Ability to debug training instabilities, loss issues, and performance bottlenecks

Bonus Skills:

Experience with custom architectures

Prior contributions to ML open-source projects

Ability to profile and optimize CUDA workloads

Research background in LLMs or generative models

Compensation & Recognition
$200 for winning competition

$400 upon completion of MageV1

Public credit as the Official Trainer of MAGIC and MageV1 on:

GitHub

HuggingFace

Documentation

Opportunity for expanded future paid roles as MAGIC evolves

Priority consideration for extended research collaborations

### **How to Apply**

1. Enter the competition:
[https://www.freelancer.com/contest/Training-MAGIC-AI-2663043/](https://www.freelancer.com/contest/Training-MAGIC-AI-2663043/)
2. Train and submit the strongest MAGIC Base model.
3. The winner receives the contract and begins work as MAGIC’s official trainer.