AI Engineer - Model Post Training

Wider Security LLC is hiring a part-time, fully remote AI Engineer to focus on post-training and alignment for large language models at scale. The work centers on supervised fine-tuning and safety-focused evaluation, with hands-on experience needed to support model tuning from 7B through 70B and beyond.

This role is built for an async-first team and emphasizes measurable outcomes across training, calibration, and adversarial robustness. You will help lead end-to-end post-training workflows, from data preparation through evaluation and deployment-oriented integration.

What you’ll do

• Lead and contribute to post-training workflows including supervised fine-tuning, instruction tuning, DPO, RLHF, RLAIF, and related alignment techniques
• Use QLoRA and other efficient fine-tuning methods across models spanning 7B to 70B+ parameter ranges
• Train models to produce reliable structured outputs under adversarial input conditions
• Build and run evaluation pipelines for safety-critical behavior, including adversarial test suites, red-team integration (for example, Garak), and regression tracking across model versions
• Calibrate decision thresholds to support tiered policy configurations, including logprob-based confidence calibration at the serving layer
• Design training approaches that preserve inference-time policy specification, enabling behavior changes without retraining
• Curate and prepare training data, evaluation sets, and preference data pipelines
• Iterate on training strategy to improve task performance, calibration, and adversarial robustness
• Document approaches and decisions clearly for a distributed, async-first team

What you’ll need

• US citizenship and current residency in the United States (firm requirement)
• Concrete, verifiable production experience post-training open-weight LLMs, with experience at 7-8B, 13-30B, or 30B+ scales welcomed, and larger-scale experience a plus
• Experience with modern open-weight model families such as Llama, Qwen, Mistral, or similar
• Hands-on experience with QLoRA, LoRA, and efficient fine-tuning methods for large models
• Background in alignment approaches including SFT, DPO, RLHF, RLAIF, or constitutional AI approaches
• Experience training models for reliable structured output (JSON, schema-constrained generation, function-call style outputs)
• Familiarity with serving stacks such as vLLM, TGI, or similar, plus comfort working with logprob-level model outputs
• Deep familiarity with distributed training frameworks including DeepSpeed, FSDP, Megatron-LM, or similar
• Proficiency in Python and comfort with multi-GPU, multi-node training infrastructure
• Ability to work independently and manage time effectively in a part-time, async-first environment

Technologies

• Python, QLoRA, LoRA, SFT, DPO, RLHF, RLAIF
• constitutional AI, vLLM, TGI, DeepSpeed, FSDP, Megatron-LM
• Llama, Qwen, Mistral, Garak

Eligibility

Applicants must be US citizens currently residing in the United States. The company is unable to consider applicants outside the US or without US citizenship, regardless of work authorization status.

Strong preferences

• Direct experience training safety classifiers, content moderation models, jailbreak or prompt injection detectors, or other trust-and-safety ML systems
• Experience with adversarial evaluation frameworks such as Garak, promptfoo, or similar
• Comfort with deployment constraints typical of regulated or restricted-network environments

Bonus qualifications

• Published research or open-source contributions related to LLM training, alignment, or AI safety
• Prior work at an AI lab, a foundation model team, or on a production safety classifier
• Experience designing or operating tiered policy systems where model behavior can be modulated at inference time

Location: Newport, RI (remote)
Job type: part time
Pay: USD 100 - 250 per hour