I turn a company's raw exports into a dataset that can train an AI model. Spreadsheets, chat logs, tickets, scans, product photos, and warehouse dumps rarely arrive as one table with a stable schema. AI Dataset Engineering is the work before fine-tuning and before a model goes into production: collect what you already have, clean it, remove duplicates and junk, label what the model must learn, and hand over files a training job can read.

The offer is concrete. Raw company data becomes a dataset fit for training AI models, with a written schema, a reproducible script, a train / validation / test split, and a quality report. I do not sell a mystery folder of CSV files. JSONL, Parquet, and WebDataset are delivery formats. The product is data whose meaning, source license, and leakage risk are known before the first training run.

This page is data preparation, not model deployment. If the corpus is already clean and you need a chat assistant, RAG, or document recognition in production, that belongs to AI implementation. Labeling ten thousand images and wiring an API are different estimates, and I do not mix them.

Scope covers sources, cleaning, deduplication, junk filters, annotation guidelines, synthetic data where real examples are scarce, and a split that keeps evaluation honest. Buying third-party databases, a legal opinion on personal data, and training the model itself stay out of scope unless the brief says otherwise. A focused pilot on one source and one task is often 3–6 weeks. Several sources plus image annotation are priced by volume, not by a slogan.