Senior LLM/VLM Training Data Engineer

Salary

$140,000–$185,000 annually

About Akoncagua AI

Akoncagua AI builds domain-specialized AI systems for technically demanding industries. Our work spans large language models, vision-language models, agentic AI, model adaptation, synthetic data, evaluation, RAG, and production AI infrastructure.

We are looking for a Senior AI Training Data Engineer with 7+ years of experience to build the datasets and data pipelines used to train domain-specialized LLMs and VLMs.

This role is focused on transforming large collections of proprietary documents, images, technical drawings, structured data, and domain knowledge into high-quality datasets for continued pretraining (CPT), supervised fine-tuning (SFT), multimodal post-training, synthetic data, and model evaluation.

What You’ll Do

• Build scalable pipelines for preparing CPT, SFT, multimodal, preference, and evaluation datasets.

• Transform PDFs, technical documents, images, OCR, tables, drawings, and structured data into model-ready training examples.

• Design domain-specific SFT task taxonomies, schemas, instruction-response formats, and dataset quality standards.

• Use frontier LLMs/VLMs as teacher models to generate synthetic training examples from proprietary domain data.

• Build automated pipelines for grounding, validating, filtering, deduplicating, and scoring synthetic and human-generated training data.

• Develop multimodal datasets combining images, OCR text, layout regions, bounding boxes, tables, semantic descriptions, and document context.

• Build text-only, image-text, and interleaved datasets for domain-specific continued pretraining and long-context model training.

• Design human-in-the-loop workflows where domain experts review, correct, and validate high-value training examples.

• Establish Gold/Silver dataset tiers, provenance, lineage, versioning, reproducibility, and auditability.

• Build held-out benchmarks and evaluation datasets while preventing train/test contamination and data leakage.

• Analyze model failures and convert them into new SFT examples, hard cases, synthetic datasets, and evaluation scenarios.

• Work closely with ML researchers and AI engineers to continuously improve model performance through better training data.

Required Qualifications

• 7+ years of experience in Data Engineering, ML Data Infrastructure, ML Engineering, AI Training Data, or a related field.

• Expert-level Python with strong software-engineering and data-processing fundamentals.

• Hands-on experience preparing datasets for LLM/VLM training, fine-tuning, post-training, or evaluation.

• Experience creating or curating SFT/instruction-response datasets.

• Strong experience with synthetic-data generation using LLMs or VLMs and automated quality validation.

• Experience processing large-scale PDF, OCR, image, table, document-layout, and multimodal datasets.

• Strong understanding of tokenization, context-length management, dataset mixing, deduplication, sampling, and train/validation/test separation.

• Experience with dataset provenance, lineage, versioning, quality metrics, and reproducible data pipelines.

• Experience with distributed data processing and cloud infrastructure, preferably AWS.

• Strong understanding of model evaluation, benchmark construction, and failure-driven dataset improvement.

Preferred Qualifications

• Experience with continued pretraining (CPT/DAPT), SFT, LoRA/QLoRA, DPO/RLHF, or other LLM post-training techniques.

• Experience preparing training data for multimodal models such as Qwen-VL/Qwen3-VL, InternVL, LLaVA, or similar VLMs.

• Experience constructing long-context datasets from 32K to 256K+ tokens.

• Experience with multimodal document understanding, OCR, visual grounding, and technical-document intelligence.

• Experience building human/expert annotation, active-learning, and model-in-the-loop data-curation systems.

• Experience creating golden datasets and domain-specific evaluation benchmarks.

• Experience working with highly technical datasets such as engineering, scientific, geospatial, CAD, financial, legal, or medical documents.