Machine Learning Engineer II (Model Training)

The Bison Group LLC•El Paso, TX

16h

About The Position

Bison Group is seeking a Machine Learning Engineer to train, fine-tune, and evaluate models for AI analysis tools. The role owns the model-training lifecycle, from dataset preparation through evaluation and deployment, and adapts models to domain-specific language and tasks. Hands-on model training experience is required; experience limited to hosted APIs or prompt engineering does not meet this requirement. The engineer will train and fine-tune open-weight language models, build and curate datasets, adapt retrieval and speech models, and develop evaluation and regression-testing workflows. The role also involves quantizing, packaging, and serving models on on-premises GPUs and constrained hardware, balancing model quality with memory use, speed, and throughput. The engineer will create reproducible training workflows and collaborate with product, UX, analysts, and subject matter experts, while documenting work so teammates can retrain, evaluate, and maintain models. Applicants must be U.S. citizens and eligible to obtain and maintain a U.S. security clearance. The Bison Group offers medical, dental, vision, life, and disability insurance, as well as accrued PTO including vacation, sick leave, and holidays. Compensation is described as a market-based range, but no specific amount is provided.

Requirements

At least 3 years of hands-on neural network training or fine-tuning experience.
Ability to explain a project end to end, including data preparation, training decisions, evaluation results, and deployment.
Hands-on fine-tuning of a 7B+ parameter open-weight language model using supervised fine-tuning (SFT), LoRA/QLoRA, or full fine-tuning, with PyTorch and Hugging Face Transformers/PEFT, NeMo, or equivalent tools.
Dataset construction skills, including cleaning, deduplication, formatting, and labeling of text data, and judgment about whether data is suitable for training.
Experience designing task-specific metrics and test sets, and diagnosing overfitting, catastrophic forgetting, and data leakage.
GPU training fundamentals, including mixed precision, gradient accumulation, checkpointing, memory budgeting, and multi-GPU basics.
Strong Python and Linux skills, including command-line work, CUDA environment management, and dependency management without internet access.
U.S. citizenship and eligibility to obtain and maintain a U.S. security clearance.

Nice To Haves

iOS development, especially on-device machine learning with Core ML or MLX.
Training small language models or distilling larger models for edge and on-device use.
Experience with NVIDIA NeMo or model serving with NIM/TensorRT-LLM.
Preference tuning with DPO, ORPO, or RLHF, or experience with reward modeling.
Training embedding or reranker models for RAG applications.
Speech recognition fine-tuning, speaker diarization, or voice activity detection.
Quantization and inference tooling such as GGUF, llama.cpp, vLLM, AWQ/GPTQ, or MLX.
Experience with ARM64 or other unified-memory platforms.
Work in defense, intelligence, or air-gapped environments; an active Secret clearance or higher.
Familiarity with qualitative research data, including focus groups, surveys, and interviews.

Responsibilities

Train and fine-tune open-weight language models using supervised fine-tuning and parameter-efficient methods such as LoRA/QLoRA; improve performance on domain terminology, task instructions, and structured outputs.
Build instruction and preference datasets from transcripts, surveys, and reference materials.
Clean, deduplicate, and label data; generate and validate synthetic examples with subject matter experts.
Train or fine-tune embedding and reranker models for retrieval-augmented generation (RAG).
Measure retrieval quality for specialized terminology and acronyms.
Adapt speech recognition, speaker diarization, and voice activity detection models for noisy audio, varied accents, and conversations with multiple speakers.
Build offline evaluation suites and regression tests covering citation and attribution accuracy, hallucination rate, and latency.
Demonstrate improvements against established baselines.
Quantize, export, and serve trained models on on-premises GPUs and constrained hardware.
Balance model quality against memory use, inference speed, and throughput.
Create reproducible training workflows with experiment tracking, versioned datasets, checkpoints, and model cards.
Package models and dependencies for deployment in environments without internet access.
Partner with product, UX, analysts, and subject matter experts to define useful model behavior.
Document experiments and decisions so teammates can retrain, evaluate, and maintain models.