Wider Security LLC is hiring for a part-time, fully remote AI Engineer focused on post-training and large language model alignment at scale. This role combines hands-on training work (from supervised fine-tuning through DPO, RLHF, and RLAIF) with safety-oriented evaluation, calibration, and structured output reliability. You will work in an async-first environment and help improve model behavior using practical, production-minded pipelines.
What you’ll work on
- 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 the 7B to 70B+ parameter range
- Train models to generate reliable structured outputs (including under adversarial input conditions)
- Build and operate evaluation pipelines for safety-critical model behavior, including adversarial test suites, red-team integration (for example, Garak), and regression tracking across model versions
- Calibrate decision thresholds against 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 an async-first team
Required qualifications
- US citizenship and currently residing 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 all welcome, and larger-scale work a plus
- Hands-on experience with modern open-weight model families such as Llama, Qwen, Mistral, or similar
- Hands-on experience with QLoRA, LoRA, and efficient fine-tuning techniques for large models
- Experience applying alignment methods including SFT, DPO, RLHF, RLAIF, or constitutional AI approaches
- Experience training for reliable structured output such as JSON, schema-constrained generation, or function-call style outputs
- Familiarity with serving stacks such as vLLM, TGI, or similar, and comfort working with logprob-level model outputs
- Deep familiarity with distributed training frameworks such as DeepSpeed, FSDP, Megatron-LM, or similar
- Proficiency in Python and comfort with multi-GPU, multi-node training infrastructure
- Ability to work independently and manage your time in a part-time, async-first environment
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 or promptfoo (or similar)
- Comfort with deployment constraints typical of regulated or restricted-network environments
Bonus qualifications
- Published research or open-source contributions in 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
Tech you’ll use
Python, QLoRA, LoRA, SFT, DPO, RLHF, RLAIF, constitutional AI, vLLM, TGI, DeepSpeed, FSDP, Megatron-LM, Llama, Qwen, Mistral, Garak
Eligibility requirement
Applicants must be US citizens currently residing in the United States. Wider Security LLC is unable to consider applicants based outside the US or those without US citizenship, regardless of work authorization status.
Location: Newport, RI (remote) • Job type: part time • Compensation: USD 100 - 250 per hourly
