We built Kvasir, a system for parallel agents to analyze data, run models, and quickly iterate on experiments based on context graphs that track data lineage. 

We built it as ML engineers who felt existing tools weren’t good enough for real-world projects we have done. Most analysis agents are notebook-centric and don’t scale beyond simple projects, and coding agents don’t understand the data. Managing experiments, runs, and iterating on results tend to be neglected.

Upload your files and give a project description like “I want to detect anomalies in this heartrate time series” or “I want to benchmark speech-to-text models from Hugging Face on this data” and parallel agents will analyze the data, generate e-charts, build processing/modeling pipelines, run experiments, and iterate on the results for as long as needed.