We kept running into the same problem with production data pipelines: when something fails, the difficult part usually isn't restarting the job. It's figuring out why it failed, what changed, and what needs to be fixed without breaking something else. We built Convalesce to investigate those failures. When a workflow fails, it collects the available failure context, traces relevant information across the connected systems, and produces a diagnosis with a proposed fix. You can then question the investigation conversationally rather than starting the investigation from scratch. For example, when a pipeline fails because of a schema or upstream-data change, the agent can correlate the pipeline failure with the relevant workflow metadata, logs, source information, and repository changes before proposing a recovery. GitHub can be connected so that a proposed code change can be opened as a PR rather than applied directly. Nothing is changed automatically in the beta; recovery actions require approval. The part we're most interested in getting feedback on is the investigation itself. We're trying to understand: Does the agent find the same context an experienced engineer would look for? Where does its root-cause analysis go wrong? Are the proposed fixes actually useful? What evidence would you require before trusting an automated recovery? It currently connects to systems including Airflow, Dagster, dbt, Spark, Snowflake, Databricks, Postgres, S3, Kafka, GitHub, and others. It's free during the beta and you connect your own infrastructure. Try it: https://convalesce.io/ Docs: https://docs.convalesce.io/docs We're happy to talk about the architecture, investigation process, or things that currently don't work well.

Convalesce

AI that diagnoses and repairs failed data workflow

AI that diagnoses and repairs failed data workflow