Machine Learning Engineer (SLM Fine - Tuning)

We're running a focused research effort to adapt small language models (SLMs) to specialized, high-value domain tasks. You'll own the full fine-tuning lifecycle, choosing base models, curating data, designing training runs, building evaluation harnesses, and shipping models that provably work. The specific domains will evolve; the constant is someone who can take any domain task and systematically turn a base SLM into a specialist for it.

Responsibilities:

• Apply parameter-efficient fine-tuning methods with a mathematically grounded understanding of low-rank adaptation, including rank selection, parameterization, scaling, and the accuracy-efficiency trade-offs of LoRA/QLoRA and related low-rank techniques.
• Choose the right adaptation strategy per task: continued pre-training (CPT), supervised fine-tuning (SFT), and RL/preference methods (DPO, GRPO, RLHF), and know when to use or sequence each.
• Curate, clean, and generate high-quality training and evaluation datasets, including synthetic data where real data is scarce.
• Design rigorous, task-specific evaluations and benchmarks based on functional, executable correctness rather than loss curves alone, with regression tracking across model versions.
• Leverage larger language models as teachers for knowledge transfer to SLMs through techniques such as knowledge distillation, structured pruning, and quantization.
• Stand up and optimize inference infrastructure (vLLM, TGI, Ollama, llama.cpp, TensorRT-LLM) across GPU, CPU, and edge targets.
• Break each target task into research milestones with clear success criteria, and drive them from baseline to production candidate.
• Stay up to date with recent research in small language models, efficient fine-tuning and adaptation strategies, model compression, and reinforcement-learning or preference-optimization methods, and translate relevant advances into reproducible experiments.
• Maintain reproducible training and evaluation pipelines with data/config versioning, checkpoints, seeds, and experiment tracking; build structured-generation and high-quality synthetic-data pipelines.

Requirements:

• 3+ years in ML engineering / applied ML, with hands-on experience fine-tuning transformer language models.
• Strong Python and PyTorch; deep fluency in the Hugging Face ecosystem (Transformers, PEFT, Datasets, Accelerate, TRL).
• Practical command of fine-tuning techniques, at minimum SFT with LoRA/QLoRA, with the judgment to reason about CPT and RL/preference methods.
• Strong understanding of base model architectures to guide model selection based on tokenizer design, pre-training distribution, licensing, and context requirements.
• Experience building training data pipelines: curation, cleaning, synthetic generation, and formatting.
• Ability to design task-specific evaluations based on functional, executable correctness rather than perplexity alone.
• Experience serving and quantizing models with at least one inference framework.
• Comfort reading recent ML literature and translating it into working code.

Nice to Have:

• Continued pre-training on domain corpora.
• RL for LLMs: reward modeling, DPO/GRPO, and verifiable-reward setups.
• Structured-output or code-generation experience (e. g., SQL/DDL), or classification fine-tuning.
• Distributed training (DeepSpeed, FSDP).
• Knowledge distillation from a larger teacher model into a smaller student.
• On-device / edge deployment.
• Open-source contributions to ML tooling or models.