$80-$220/hr machine learning engineering work, on your schedule
Review AI-generated training code, data pipelines, and deployment the way you'd review a model before it ships. Catch the label leak, the eval that overstates, the serving path that silently degrades. The judgment from running real models in production - where they fail in ways notebooks never show - is what AI labs need.
Trusted by top research companies
Rated by experts worldwide
Hi, we're Zac and Jack, the founders of Terac. We want to talk to you directly, because you are the most important part of what we're building.
Terac is a community of experts. People who have spent years getting good at something specific and hard. The world is about to need more of you, not less. As AI takes on more of the world's work, the bottleneck shifts to the people who actually know what they're talking about.
Expert labor is the rarest resource in the world right now, and it is shockingly hard to find. The companies that need an ML engineer's eye on a training pipeline that leaks labels spend weeks chasing people, paying placement fees, and settling for whoever is available. Meanwhile thousands of qualified people are sitting with knowledge that no one ever asks for.
That gap is what we're here to close. Every project that lands on Terac is routed to the people who actually know the answer, on their schedule, paid fairly, and only when the work is verified. No middleman taking a cut of your time. No vague gigs. No chasing checks.
We care about every single person in this community. If you join Terac, you're not a row in a database to us. We read the feedback. We answer the emails. We will fight for you when a customer is being unreasonable, and we will be honest with you when something on our side is broken. The quality of this panel is our entire company, and we owe you a serious bar.
If you've made it this far, here is what we're asking: claim your profile. Put your expertise on the record. Let the world's most ambitious teams come find you for the work only you can do.
ML Engineering questions
Still curious? Write to us at support@terac.com.
Niche specializations are often the most valuable. Frontier labs have well-covered generalist feedback and real gaps in areas like post-training alignment, efficient inference, and novel architecture design, so your depth is an asset rather than a limitation. You will be matched to tasks that fit your sub-domain rather than assigned a generic queue.
Tasks vary by project but typically include reviewing Python training scripts (PyTorch or JAX), evaluating the correctness of loss function implementations, critiquing explanations of gradient flow or backpropagation, and assessing whether a model card accurately describes training data and evaluation methodology. Some projects ask you to write worked examples that demonstrate expert reasoning through a problem, such as debugging a vanishing-gradient issue or designing an evaluation suite for a retrieval-augmented generation system.
Industry certifications like the Google Professional ML Engineer and AWS Machine Learning Specialty are accepted as supporting evidence of expertise. They are weighted alongside your work history and any published research or open-source contributions. A PhD is not required, and practitioners with strong industry credentials and production experience are routinely onboarded.
Some tasks do involve evaluating model behavior on sensitive topics, including fairness audits, toxicity classifier outputs, and assessments of whether a model correctly refuses harmful requests. You will not be asked to generate harmful content yourself. If a specific task conflicts with your professional ethics or employer policies, you can decline it without penalty and will be routed to other work.
MLOps and inference-infrastructure expertise is genuinely in demand. Tasks in this area include reviewing latency benchmarks, evaluating the correctness of ONNX or TensorRT export workflows, and assessing advice about serving architectures like Triton Inference Server or vLLM. You do not need a research background, and production-focused engineers often provide feedback that research-oriented reviewers miss.
Why your expertise matters
ML engineers are the exact practitioners who catch subtle modeling mistakes that general reviewers miss: a misapplied loss function, a data leakage pattern in a feature pipeline, or an incorrect claim about model calibration. Frontier models learn to reason about ML workflows from training data, and that data needs to reflect how working engineers actually debug, iterate, and make tradeoffs - not how textbooks describe them. Your judgment on when a particular architecture choice is appropriate for a given scale, latency, and data constraint is exactly what current training corpora lack.
How pay works
Rates toward the $220/hr ceiling go to engineers with deep specialization in areas currently under-represented in training data, such as production-scale MLOps, reinforcement learning from human feedback, or on-device model optimization. All work is fully remote and billed by the verified hour. Payment releases after Terac confirms the deliverable meets the scope defined for that task, so there are no invoicing delays tied to client approval cycles.
What the work looks like
A sample of the machine learning engineering work you would pick up. Every project is scoped, remote, and paid on verified completion.
- Review an AI-generated PyTorch training script and flag incorrect gradient accumulation logic that would produce silently wrong results at large batch sizes.
- Evaluate a model card drafted by a language model and correct factual errors about the training dataset composition, evaluation splits, and reported metric definitions.
- Write a worked example showing how to properly detect and remediate target leakage in a tabular feature pipeline, including the specific pandas and scikit-learn patterns an experienced engineer would use.
- Assess an AI explanation of distributed training strategies and identify where the description of ZeRO optimizer stages conflates memory sharding with gradient checkpointing.
- Create a multi-step reasoning trace that walks through diagnosing a training instability, covering loss curves, gradient norms, learning rate schedules, and the decision logic for each diagnostic step.
- Judge whether an AI-generated answer about choosing between fine-tuning and retrieval-augmented generation correctly accounts for latency constraints, data freshness requirements, and deployment environment.
Specialties we match
ML Engineering projects span a wide range of focus areas. Tell us where you go deep and we route the work that fits.
- PyTorch / JAX training loops
- Distributed training (FSDP, DeepSpeed)
- Feature engineering and data pipelines
- Model evaluation and benchmarking
- MLOps and CI/CD for ML (MLflow, Weights & Biases)
- Quantization and model compression
- RLHF and reward modeling
- On-device inference (TensorRT, ONNX, CoreML)
- Experiment tracking and reproducibility
- Hyperparameter optimization (Optuna, Ray Tune)
- Data labeling quality and annotation pipelines
- Transformer architecture internals
