We noticed that when there’s an application error it often takes too much time to figure out which issues are important and how to fix them. Because the overhead is too high, people (including us…) sometimes just pay attention to bugs that come directly from a user complaint.

We were initially building an AI coding assistant that traced code execution to teach it how a specific codebase behaves (see a demo here: https://youtu.be/2LjVH2sH2Q8). In doing so, we realized that if you capture the right code-level runtime data (local variables, request, arguments, etc.) LLMs become quite effective at figuring out the impact of errors and debugging them.

From that insight, we began working on Decipher which monitors for errors and collects all the relevant context in one place. This includes logs, endpoint, request bodies, the stack trace, local variables, and relevant code.

With this centralized context, we use LLMs to summarize the issue and investigate a solution. What’s cool is that by summarizing this stuff, you can get a gist of the bug directly in a Slack alert (rolling out soon).

For example, here’s an issue on a shopping app with a Flask backend, there was a `List Index out of Range error`: https://imgur.com/a/uEfoS9k) You can see all the relevant details are brought to the forefront and the AI was able to summarize the issue and hint at what the solution might be.

We’re exploring capturing additional context like the function call graph and argument values and analyzing session replays to figure out the user impact of errors.

Would love for folks to try out our free beta and would really appreciate any feedback! We currently support Express, Next.js, and Flask with just a few minute setup (install the library here: prod.getdecipher.com/docs)

Library code: https://github.com/decipherai/decipher-client-js, https://github.com/decipherai/decipher-client-py