ML Engineer

What you'll do

This is hands-on engineering work. You create, solve, review and validate difficult machine learning engineering tasks. A representative one implements or modifies a model, builds a reproducible training or inference workflow, optimises memory or throughput, debugs a numerical or system-level failure, and then proves the result meets objective correctness and performance criteria.

In practice that spans model components, data pipelines, evaluation systems and numerical methods; tensor operations, automatic differentiation, architectures, tokenisation, batching and generation; latency, throughput, memory use and hardware utilisation. The debugging side covers numerical instability, incorrect tensor behaviour, memory bottlenecks, distributed system failures and performance regressions. You also review AI-generated code and technical solutions for correctness, efficiency and engineering quality, design the tests and benchmarks that verify them, and document the trade-offs and limitations.

The ad says outright that it wants people who understand the systems beneath the high-level APIs and can explain what they personally built, optimised or operated. Read that as a screening statement: the AI interview stage exists to probe exactly that.

What you need

• A master's degree or PhD in computer science, machine learning, AI, applied mathematics, statistics, engineering or a closely related quantitative discipline
• Strong professional or research experience in machine learning
• Practical proficiency with Python
• Meaningful experience with at least two relevant frameworks, libraries or inference tools. The ad names PyTorch, JAX, NumPy and SciPy, SGLang, vLLM, llama.cpp, Hugging Face Transformers and Hugging Face Tokenizers, and says equivalents may count where you can show directly relevant depth
• A strong understanding of model training, evaluation, numerical computation or inference
• The ability to debug ML systems below surface-level API usage
• The ability to explain implementation decisions, performance trade-offs and failure modes clearly
• Experience building reproducible technical workflows

Experience at a well-established technology company, AI laboratory, research organisation or comparable engineering environment is strongly preferred, with exceptional open-source or academic work offered as the alternative route.

Worth flagging: this is one of the few micro1 listings where the degree sits under required qualifications rather than preferred, and no equivalent-experience substitute is published for it.

How hiring works

Four stages, published on the ad: apply and complete the screening questions, sit an AI interview of roughly 30 minutes, complete a technical assessment if one is required, then a hiring manager review.

micro1 says it typically fills these roles within 48 hours and expects a selected expert to start their first task within 24 to 48 hours of finishing onboarding. Start the application when you are genuinely free to begin, not as a placeholder.

What it pays

$100–150/hour, and unusually for this board the compensation structure is published rather than left blank: pay is output-based, per task that meets the project specifications, and minimum submission requirements apply.

That changes how to read the band. The hourly figure is an expected equivalent, not a guarantee of paid time, and how close you land to it depends on how quickly you turn out work that passes specification. Engineers who are fast on this kind of task often clear the implied rate; careful slow work on rejected submissions is unpaid. What the ad does not publish is what the minimum submission requirements actually are, which is the first thing to ask.

The engagement is roughly 15 hours a week, with flexible hours and days including weekends if you want them.

Worth knowing

Good:

• $100–150/hour with the pay mechanism actually stated, which most listings on this board omit
• 100 openings, the largest count here by a wide margin
• Global and fully remote, with no country restriction stated
• Engineering on real models, with optimisation and debugging, and reviewing generated code as only one part
• Roughly 15 hours a week on hours you pick, which fits alongside a full-time job
• Exceptional open-source or academic depth can substitute for a brand-name employer

Less good:

• A master's or PhD is a required qualification, with no equivalent-experience route published
• Output-based pay means time spent on submissions that miss specification is unpaid
• "Minimum submission requirements apply" with no figure attached
• A 48-hour hiring cycle and a 24 to 48 hour start expectation, so it suits people who are free now
• A 30-minute AI interview as a screening stage, which some engineers decline on principle
• Contractor engagement: no benefits, no notice period, your own tax to manage
• The customer project and the lab behind it are not named

About this listing

Posted by micro1 on its own jobs portal, confirmed open on 12 September 2026. The $100–150/hour band, the 100 openings, the roughly 15 hours a week, the four-stage process, the output-based compensation and every qualification above are its own published text for this listing.

Pay mechanics differ from listing to listing at micro1, so do not carry assumptions across from its hourly evaluation roles: this one publishes per-task, output-based pay and they do not. The minimum submission requirement is not quantified anywhere on the ad, and the customer is not named. Screening and payment run at the platform level. See micro1.