Machine Learning Engineer

What it does

Machine Learning Engineer is a Cursor agent playbook named ml-engineer. AgenticKit installs it so a session can adopt this role instead of acting as a generic coding assistant.

The ml-engineer agent helps solo founders adapt open-source or proprietary AI models to custom domain data. It structures training datasets (JSONL), configures fine-tuning jobs (LoRA / OpenAI), and evaluates model quality against test benchmarks.

Prepares dataset fine-tuning, model evaluation, and Python ML pipelines. Machine learning engineer for Cursor providing dataset formatting, fine-tuning scripts, model evaluation metrics, and inference optimization.

Why it exists

A pile of agent files is not a team. ml-engineer exists so one job, machine learning engineer, has a written brief, required skills, and a stop condition.

It ships in the Engineering Kit. Skills it must read: Machine Learning Fine-Tuning & Model Evaluation; LLM Integration & Streaming AI Workflows; System Prompt Engineering & Few-Shot Design. Commands that adopt this role: /ml-finetune: ML Model Fine-Tuning & Dataset Pipeline; /ai-feature: AI Integration & Vector Pipeline Builder; /rag-pipeline: RAG Pipeline & Document Ingestion System.

When to use it

• Fine-tuning models on domain-specific proprietary data
• Evaluating prompt accuracy and model benchmarking
• Building Python data preparation and training scripts

When not to use it

• Do not ask ml-engineer to do a different role's job. If you need a launch post, switch agents.
• Do not keep the same chat after this agent has finished its artifact. Start a reviewer in a new thread.
• Do not invoke every agent in the kit for a small change.

Example workflow

1. Install the Engineering Kit so .cursor/agents/ml-engineer is on disk.
2. Run /ml-finetune: ML Model Fine-Tuning & Dataset Pipeline, or start a chat and tell Cursor to adopt the ml-engineer role.
3. The agent should read: Machine Learning Fine-Tuning & Model Evaluation; LLM Integration & Streaming AI Workflows; System Prompt Engineering & Few-Shot Design.
4. It produces the artifact for this role only, then stops.
5. A different agent or you review. Same-chat self-review is not a review.

Example usage

Example: "You are ml-engineer. Fine-tuning models on domain-specific proprietary data. Read ml-training before you edit."

Or let the pipeline invoke it: /ml-finetune: ML Model Fine-Tuning & Dataset Pipeline.

Capabilities you should actually see: JSONL dataset formatting, validation, and deduplication

Example output

• ml-engineer should leave files or a written verdict, not a vibe check. JSONL dataset formatting, validation, and deduplication LoRA and full fine-tuning pipelines for Llama / Mistral models
• Engineering agents should touch the slice they were given (schema, route, test, or review note) and nothing else.

Best practices

• JSONL dataset formatting, validation, and deduplication
• LoRA and full fine-tuning pipelines for Llama / Mistral models
• Model inference optimization and latency reduction
• Automated evaluation suites for hallucination and accuracy
• One role per chat unless a command is explicitly orchestrating a sequence.

Common mistakes

• Using ml-engineer as a synonym for 'the Cursor agent'. It is a brief, not the product.
• Skipping the required skills and hoping the role name is enough.
• Letting the writer approve its own PR.
• Invoking this agent and three unrelated ones in the same prompt.