May 18, 2024
Summarization header image

As somebody who takes loads of notes, I’m all the time looking out for instruments and techniques that may assist me to refine my very own note-taking course of (such because the Cornell Technique). And whereas I usually want pen and paper (as a result of it’s proven to assist with retention and synthesis), there’s no denying that know-how may also help to boost our built-up talents. That is very true in conditions comparable to conferences, the place actively taking part and taking notes on the similar time might be in battle with each other. The distraction of trying right down to jot down notes or tapping away on the keyboard could make it exhausting to remain engaged within the dialog, because it forces us to make fast choices about what particulars are essential, and there’s all the time the chance of lacking essential particulars whereas attempting to seize earlier ones. To not point out, when confronted with back-to-back-to-back conferences, the problem of summarizing and extracting essential particulars from pages of notes is compounding – and when thought of at a gaggle degree, there may be significant individual and group time waste in fashionable enterprise with most of these administrative overhead.

Confronted with these issues every day, my staff – a small tiger staff I wish to name OCTO (Workplace of the CTO) – noticed a chance to make use of AI to enhance our staff conferences. They’ve developed a easy, and easy proof of idea for ourselves, that makes use of AWS companies like Lambda, Transcribe, and Bedrock to transcribe and summarize our digital staff conferences. It permits us to collect notes from our conferences, however keep centered on the dialog itself, because the granular particulars of the dialogue are mechanically captured (it even creates a listing of to-dos). And right this moment, we’re open sourcing the device, which our staff calls “Distill”, within the hopes that others would possibly discover this handy as properly: https://github.com/aws-samples/amazon-bedrock-audio-summarizer.

On this publish, I’ll stroll you thru the high-level structure of our mission, the way it works, and offer you a preview of how I’ve been working alongside Amazon Q Developer to show Distill right into a Rust CLI.

The anatomy of a easy audio summarization app

The app itself is simple — and this is intentional. I subscribe to the idea that systems should be made as simple as possible, but no simpler. First, we upload an audio file of our meeting to an S3 bucket. Then an S3 trigger notifies a Lambda function, which initiates the transcription process. An Event Bridge rule is used to automatically invoke a second Lambda function when any Transcribe job beginning with summarizer- has a newly updated status of COMPLETED. Once the transcription is complete, this Lambda function takes the transcript and sends it with an instruction prompt to Bedrock to create a summary. In our case, we’re using Claude 3 Sonnet for inference, but you can adapt the code to use any model available to you in Bedrock. When inference is complete, the summary of our meeting — including high-level takeaways and any to-dos — is stored back in our S3 bucket.

Distill architecture diagram

I’ve spoken many times about the importance of treating infrastructure as code, and as such, we’ve used the AWS CDK to manage this project’s infrastructure. The CDK gives us a reliable, consistent way to deploy resources, and ensure that infrastructure is sharable to anyone. Beyond that, it also gave us a good way to rapidly iterate on our ideas.

Using Distill

If you try this (and I hope that you will), the setup is quick. Clone the repo, and observe the steps within the README to deploy the app infrastructure to your account utilizing the CDK. After that, there are two methods to make use of the device:

  1. Drop an audio file immediately into the supply folder of the S3 bucket created for you, wait a couple of minutes, then view the leads to the processed folder.
  2. Use the Jupyter pocket book we put collectively to step by way of the method of importing audio, monitoring the transcription, and retrieving the audio abstract.

Right here’s an instance output (minimally sanitized) from a current OCTO staff assembly that solely a part of the staff was capable of attend:

Here’s a abstract of the dialog in readable paragraphs:

The group mentioned potential content material concepts and approaches for upcoming occasions like VivaTech, and re:Invent. There have been solutions round keynotes versus having fireplace chats or panel discussions. The significance of crafting thought-provoking upcoming occasions was emphasised.

Recapping Werner’s current Asia tour, the staff mirrored on the highlights like partaking with native college college students, builders, startups, and underserved communities. Indonesia’s initiatives round incapacity inclusion have been praised. Helpful suggestions was shared on logistics, balancing work with downtime, and optimum occasion codecs for Werner. The group plans to research turning these learnings into an inner publication.

Different subjects lined included upcoming advisory conferences, which Jeff could attend nearly, and the evolving position of the fashionable CTO with elevated concentrate on social impression and world views.

Key motion objects:

  • Reschedule staff assembly to subsequent week
  • Lisa to flow into upcoming advisory assembly agenda when out there
  • Roger to draft potential panel questions for VivaTech
  • Discover recording/streaming choices for VivaTech panel
  • Decide content material possession between groups for summarizing Asia tour highlights

What’s extra, the staff has created a Slack webhook that mechanically posts these summaries to a staff channel, in order that those that couldn’t attend can make amends for what was mentioned and shortly evaluate motion objects.

Bear in mind, AI is just not good. A few of the summaries we get again, the above included, have errors that want guide adjustment. However that’s okay, as a result of it nonetheless quickens our processes. It’s merely a reminder that we should nonetheless be discerning and concerned within the course of. Crucial pondering is as essential now because it has ever been.

There’s worth in chipping away at on a regular basis issues

This is just one example of a simple app that can be built quickly, deployed in the cloud, and lead to organizational efficiencies. Depending on which study you look at, around 30% of corporate employees say that they don’t complete their action items because they can’t remember key information from meetings. We can start to chip away at stats like that by having tailored notes delivered to you immediately after a meeting, or an assistant that automatically creates work items from a meeting and assigns them to the right person. It’s not always about solving the “big” problem in one swoop with technology. Sometimes it’s about chipping away at everyday problems. Finding simple solutions that become the foundation for incremental and meaningful innovation.

I’m particularly interested in where this goes next. We now live in a world where an AI powered bot can sit on your calls and can act in real time. Taking notes, answering questions, tracking tasks, removing PII, even looking things up that would have otherwise been distracting and slowing down the call while one individual tried to find the data. By sharing our simple app, the intention isn’t to show off “something shiny and new”, it’s to show you that if we can build it, so can you. And I’m curious to see how the open-source community will use it. How they’ll extend it. What they’ll create on top of it. And this is what I find really exciting — the potential for simple AI-based tools to help us in more and more ways. Not as replacements for human ingenuity, but aides that make us better.

To that end, working on this project with my team has inspired me to take on my own pet project: turning this tool into a Rust CLI.

Building a Rust CLI from scratch

I blame Marc Brooker and Colm MacCárthaigh for turning me right into a Rust fanatic. I’m a techniques programmer at coronary heart, and that coronary heart began to beat loads sooner the extra acquainted I obtained with the language. And it grew to become much more essential to me after coming throughout Rui Pereira’s wonderful research on the power, time, and reminiscence consumption of various programming languages, once I realized it’s super potential to assist us construct extra sustainably within the cloud.

Throughout our experiments with Distill, we wished to see what impact shifting a operate from Python to Rust would appear to be. With the CDK, it was simple to make a fast change to our stack that permit us transfer a Lambda operate to the AL2023 runtime, then deploy a Rust-based model of the code. Should you’re curious, the operate averaged chilly begins that have been 12x sooner (34ms vs 410ms) and used 73% much less reminiscence (21MB vs 79MB) than its Python variant. Impressed, I made a decision to essentially get my palms soiled. I used to be going to show this mission right into a command line utility, and put a few of what I’ve discovered in Ken Youens-Clark’s “Command Line Rust” into apply.

I’ve all the time beloved working from the command line. Each grep, cat, and curl into that little black field jogs my memory a number of driving an outdated automobile. It might be slightly bit tougher to show, it’d make some noises and complain, however you’re feeling a connection to the machine. And being energetic with the code, very similar to taking notes, helps issues stick.

Not being a Rust guru, I made a decision to place Q to the take a look at. I nonetheless have loads of questions in regards to the language, idioms, the possession mannequin, and customary libraries I’d seen in pattern code, like Tokio. If I’m being trustworthy, studying interpret what the compiler is objecting to might be the toughest half for me of programming in Rust. With Q open in my IDE, it was simple to fireplace off “silly” questions with out stigma, and utilizing the references it supplied meant that I didn’t must dig by way of troves of documentation.

Summary of Tokio

Because the CLI began to take form, Q performed a extra important position, offering deeper insights that knowledgeable coding and design choices. As an example, I used to be curious whether or not utilizing slice references would introduce inefficiencies with massive lists of things. Q promptly defined that whereas slices of arrays could possibly be extra environment friendly than creating new arrays, there’s a chance of efficiency impacts at scale. It felt like a dialog – I may bounce concepts off of Q, freely ask observe up questions, and obtain quick, non-judgmental responses.

Advice from Q on slices in Rust

The very last thing I’ll point out is the characteristic to ship code on to Q. I’ve been experimenting with code refactoring and optimization, and it has helped me construct a greater understanding of Rust, and pushed me to assume extra critically in regards to the code I’ve written. It goes to point out simply how essential it’s to create instruments that meet builders the place they’re already snug — in my case, the IDE.

Send code to Q

Coming quickly…

Within the subsequent few weeks, the plan is to share my code for my Rust CLI. I would like a little bit of time to shine this off, and have people with a bit extra expertise evaluate it, however right here’s a sneak peek:

Sneak peak of the Rust CLI

As all the time, now go construct! And get your palms soiled whereas doing it.