ASSISTANT
A Multi-Agent AI System, Not a Chatbot
Specialist routing, 50+ callable tools, and guardrails that prevent hallucinations by design — built for the complexity of real data stacks.
9 Purpose-Built Modes
Ask, Debug, Plan, Create, Fix, Explain, Suggest, Bundle, and Onboard — every mode is tuned for a specific workflow.
SQL Specialist
Dialect-aware queries for BigQuery, Snowflake, Postgres, Redshift, and Databricks. Zero "wrong dialect" answers.
Logs Specialist
Parses Airflow, Dagster, and Prefect logs automatically for root-cause, with structured failure classification.
CodeFix Specialist
Autonomous repair of SQL, Python, and dbt — with downstream-aware patching and one-click GitHub PRs.
50+ Agentic Tools
Warehouse, lineage, coverage, observability, change plans, dbt, Airflow, Dagster, Prefect, n8n, Fivetran, Airbyte, Looker, Tableau, Power BI, Metabase, GitHub.
Weighted Intent Detection
Mode is chosen from keywords (40%), question form (20%), history (20%), and target (20%) — not brittle keyword match.
Context Builder
Pre-fetches schemas and ranks tables by relevance — exact match +10, partial +5, column +3 — so answers stay grounded.
Streaming Thinking Steps
Watch the AI reason in real time. Every tool call, evidence fetch, and synthesis step is visible as it happens.
Tool Confirmation UI
Every write operation is gated by an explicit approval step. Agentic, but never silent.
Smart @Mentions
Autocomplete for tables, models, sources, and connectors. Reference anything in your stack inline.
Evidence → Specialist → Synthesis
Iterative reasoning with graceful fallback chains — so a single failed tool never kills the response.
No Hallucinations by Design
Structured workflows enforce lineage checks before proposing anything. Fix mode requires impact analysis first.
One Assistant, Nine Workflows
Change Plans — Safe Autonomy
The assistant can act on production — but only through a structured plan with validation layers and rollback built in.
Lineage Analysis First
Fix mode mandates full lineage impact analysis before proposing any change. No blind patches.
Downstream Auto-Patching
When a fix breaks downstream objects, the assistant patches every affected model automatically.
Multi-Layer Validation
Static syntax → CI compile/tests → sandbox shadow run. Every change passes through all three.
Rollback Strategies
Every plan ships with a rollback path attached — revert in one click if anything drifts.
One-Click GitHub PRs
Create a PR with the full plan, validation status, and rollback plan attached automatically.
Shareable Plan Modals
Send a review link to your team — they see the plan, evidence, and validation before approving.