SignalPilot is a data development lifecycle management platform that helps data teams and AI agents produce correct, evidence-backed answers from enterprise warehouses. It combines maintained context, automated evaluations, and a governed cloud sandbox to catch plausible-but-wrong SQL before it reaches production dashboards.
Funding
Funding not disclosed
Founders
Product
Problem
Data teams and AI agents frequently produce SQL that runs successfully but returns incorrect results, such as double-counted revenue from invoice line-item joins or misread cumulative totals. These plausible errors surface weeks later in leadership meetings, forcing costly rework and eroding trust in data-driven decisions.
Solution
SignalPilot is a development lifecycle management platform that maintains context, runs evaluations on every change, and provides a governed cloud sandbox where agents execute against real warehouse schemas. The system reasons about data grain rather than column names, applies deterministic checks like row counts and fan-out ratios before anything is called done, and builds an evidence graph linking every answer to its source queries. It integrates with dbt projects and Markdown definitions, supports agents from Claude Cowork, Codex, or Slack, and flags duplication or cumulative-value errors automatically. The platform is ranked #1 on Spider 2.0-DBT and ADE-Bench, with public transcripts of every benchmark run.
Target Audience
Primary customers are data engineering and analytics teams at mid-market and enterprise companies who own revenue-critical data models and need to deploy AI agents safely against production warehouses.
Features
- Maintained context layer that stores data definitions, business logic, and knowledge base in dbt, Apache Ossie, and Markdown within the user's repo
- Automated eval suite with 30 scheduled runs per month, covering row counts, fan-out ratios, cardinality, column completeness, and value spot-checks
- Governed cloud sandbox that runs queries against Snowflake, Databricks, BigQuery, Postgres, DuckDB, and Redshift with optional self-hosting via Docker Compose
- Evidence graph on every answer, tracing results back to source queries and warehouse rows
- Verifier agents that catch plausible-but-wrong SQL patterns, such as invoice line-item duplication and cumulative column misreads
- Integration with Claude Cowork, Codex, and Slack for agent access, plus FHIR-compatible APIs for EHR export where applicable
- Public benchmark transcripts and task-by-task evaluation results on Spider 2.0-DBT and ADE-Bench