This essay really made me rethink the Prisoner's Dilemma - the argument that cooperation can be the optimal strategy when considering dependencies and correlations between players is refreshing. Enjoyed hearing a challenge to the traditional assumption of selfish defection

We built Summ, a tool that provides intelligent search and question-answering across large sets of transcripts. Summs turn your unstructured transcripts into queryable SQL and JSON!

Yasyf and I are both very active users of call note-taking tools like Otter. While doing a recent deep-dive on finance automation, we realized that we had built up a large corpus of user interview transcripts, but were struggling to pattern-match across them. Getting answers to simple questions like "What are the most common metrics that customers track in their HR departments?" was taking hours to aggregate. To answer that question, we would need to write a handful of search queries, comb through transcripts to pull in context, copy the results into a doc, and then cobble together a final answer.

To automate this process, we built a tool that could thoroughly answer questions about a set of uploaded transcripts and filter by transcript attributes such as department, role, and company industry.

For inference, we pick the best result of three methods. The first two methods use LLMs to build structured datasets, creating and populating a "database" with relevant metrics to answer a question. This structured data is then used by the LLM to run calculations on (eg. what percent of users report using spend management features). The last method uses recursive summarization, surfacing and summarizing to eventually answer the query.