Weiser AI - Data Quality Framework

YAML-Based Configuration

Define your data quality checks and agent evals with simple, human-readable YAML. No complex code required - just describe what you want to validate and Weiser handles the rest. The declarative format is LLM-friendly, so large language models can easily read and generate Weiser configurations.

Agent Evals

Built for data-powered agents. Evaluate NL-to-SQL and BI agents with the same declarative config as your checks - compare arms, score runs, and gate releases. Weiser's own data quality checks are wired into eval scoring, so when a data incident breaks an answer, blame lands on the data, not the agent.

Enterprise-Ready Scale

Supports PostgreSQL, MySQL, Databricks, Snowflake, BigQuery, Cube, and agent frameworks like Pydantic AI and Strands (more to come), scaling to handle millions of records with advanced statistical analysis and anomaly detection.

Simple YAML Configuration

Define data quality checks with intuitive YAML syntax. Perfect for version control, team collaboration, and AI-assisted generation.

# weiser-config.yaml
checks:
  - name: orders_exist
    dataset: orders
    type: row_count
    condition: gt
    threshold: 0
    
  - name: revenue_validation
    dataset: orders
    type: sum
    measure: order_amount
    condition: ge
    threshold: 10000
    filter: status = 'completed'
    
  - name: data_completeness
    dataset: customers
    type: not_empty_pct
    dimensions: [email, phone]
    condition: le
    threshold: 0.05  # Max 5% NULL

LLM-Friendly Design

Weiser's human-readable configuration makes it perfect for AI assistance. LLMs can easily understand, generate, and modify data quality checks.

🤖 AI Code Generation

LLMs can generate Weiser configs from natural language descriptions

📝 Self-Documenting

YAML structure is inherently readable by both humans and AI

🔄 Easy Modification

AI assistants can update and refine existing configurations

💡 Smart Suggestions

LLMs can recommend new checks based on your data schema

Agent Evals

Evaluate NL-to-SQL and BI agents the same way you define checks — declaratively in YAML. Compare arms, score runs, and gate releases, all in the same config file as your data quality checks.

# evals.yaml
agent_variants:
  - name: baseline
    framework: pydantic_ai
    entrypoint: myapp.eval_agents.build_bi_agent
    tools: [list_views, describe_view, query, submit_answer]

eval_suites:
  - name: lookup_tool_ablation
    arms:
      - name: baseline
        agent_variant: baseline
        semantic_layer: local_sl
    metrics:
      - type: reference_value_match
      - type: llm_judge
        name: sql_soundness

Explore Agent Evals

Evals-As-YAML

The experiment design — tools, prompts, models, and test cases — lives entirely in YAML. Python is only needed for a one-time agent factory.

🧪 Arm Comparisons

Ablate one variable at a time across arms sharing a golden set

🔗 Data-Quality Attribution

Weiser's own DQ checks are wired into eval scoring — a data incident is blamed on the data, not the agent

⚖️ Deterministic + LLM-Judge Metrics

Zero-cost checks alongside rubric-based judging, calibrated against human labels

🚦 CI Regression Gating

Gate candidate runs against a baseline before they ship