We have taken over 2000+ project tickets in Linear from different Relevance AI teams in 1 month, then applied clustering on them to find what task categories take longer to complete and which are completed in a fast turnaround time.
Product and project management revolves heavily around software in modern businesses such as Jira, Notion, Linear & Monday, etc.
For some companies, 1k+ task tickets can be created in these project management platforms to manage these projects.
This is a wealth of unstructured data to draw insights from such as the agility, weakness and expertise of different company teams. All of this can surfaced by analyzing the unstructured text from these tickets.
This showcase is centered around demonstrating how we at Relevance AI actually utilized our own software to surface those insights for our team.
Compiled from over 2000+ tickets taken from our project management system Linear. These were then exported into an CSV file and subsequently uploaded as a dataset in the Relevance AI platform, then vectorised & clustered.
- The highest quantity of tickets were related to the development of our Clustering product.
- 76 tickets total with an average completion time of 11.7 hours per ticket.
- Out of 76 tickets total, only 38 tickets were completed in that month.
- Our fastest tickets were completed in 5.3 hours (note our team often works on tickets simultaneously). The majority of those tickets belonged to the Product team.
- Our slowest tickets were related to writing vector-based content, at an average of 13.4 hours.
- Out of 22 tickets total, only 14 have been completed in that month.
- If we filter the data within our dashboard by the “Backend” team, the most completed tickets were related to “Chunk vector search”
Two apps were created very quickly, thanks to our rapid experimentation features.
- An app that utilizes vector embeddings from OpenAI
- An app that utilizes vector embeddings from SentenceTransformers
In order, the steps performed were:
- Ingesting the exported csv from our project management tool
- Vectorizing the ticket titles
- Clustering on the vectors created from the ticket titles, ranging from 30-120 number of clusters
- Validating those cluster quantitatively using our Cluster Report card which looks at range of statistical scores such as Silhouette score, Dunn index, etc.
- Validating those clusters qualitatively by viewing the distance between words using both the cluster app and our 3d projector tool.