Generative AI for Data Teams

• How LLMs generate text
• Context windows and tokens
• Typical failure modes on data tasks
• Model families and when to use each
• Cost per token and budgeting for data workloads

🧪 Hands-on: Compare LLM answers against ground-truth queries

• Prompting for SQL generation
• Reviewing and testing generated SQL
• Auto-documenting tables and columns
• Generating dbt-style models with an LLM
• Guardrails against destructive SQL

🧪 Hands-on: Generate, test and document queries on a sample warehouse

• Schema-guided extraction
• Validating JSON output
• Handling errors at scale
• Classifying free-text fields at scale
• Batch vs real-time LLM calls

🧪 Hands-on: Extract fields from invoices into a table

• Chunking strategies
• Embeddings and vector storage
• Metadata and filters
• Choosing embedding models for tabular and text data
• Refreshing embeddings when source data changes

🧪 Hands-on: Load documents into pgvector

• Retrieval and prompt assembly
• Citations
• Basic evaluation
• Hybrid keyword and vector retrieval
• Answer quality checks with a test set

🧪 Hands-on: Build a cited Q&A over data docs

• Classifying sensitive data
• Masking and access control
• Logging and review
• Data contracts for AI consumers
• Retention and deletion of prompts and outputs

🧪 Hands-on: Add masking and access checks to the pipeline

• Generating data tests from column descriptions
• Explaining anomalies in plain language
• Triage of failed checks
• Human review workflow

🧪 Hands-on: Add LLM-written explanations to a failing data-quality report

• Scheduling LLM jobs in Airflow
• Caching and retries
• Logging prompts for audit
• Monitoring cost and drift

🧪 Hands-on: Orchestrate the extraction pipeline with retries and a cost dashboard

• Scoping with a stakeholder
• Building the pipeline end to end
• Evaluation report
• Demo and review

🧪 Hands-on: Ship a governed, cited Q&A over a real internal dataset