I’ll go back and talk about earlier parts of the project like why I wanted to do it, but since our tech stack is starting to stick (for now) I wanna mark the milestone.
What the project is though, feels like as good a place as any to start. My hypothesis, I happen to think that there’s better ways to get information than using social media as your hub. Since I’ve dabbled in AI as a practitioner for years, generative AI felt like the point to dive in and see what it can really do.
Up until the start of the pandemic, I usually had some type of tech project going on. Some of them started out great, but life does life things and progress gets slow. You drift toward new opportunities but I’ve found the team you pull together along the way walks away with new experience and insights.
I figured I’d pick that energy back up while I upskill and try to find a problem to play with. In this case: how difficult is it for an LLM to create 20-30 minute briefings on the latest news at a scheduled frequency. And could one extend capability that to create custom “personas” for users to adjust their content preference weights toward certain content like sports.
So i picked the tech stack I’m most familiar with (AWS) and started to build around it. Keep in mind, this will certainly evolve over time. This post just happens to be the one where I chose a pattern for the audio content. I thought about it like a podcast when it started but the people who have listened to an episode or two think it feels more like a newscast and I agree.
I’m also using AI as an assistant, mostly to investigate fixes in my pipelines and workflows. What I’m discovering about the development experience is that you still need to understand the software development process to multiply your efficiency.
I’m also discovering I’m going to also spend a lot of time cleaning up technical debt. AI doesn’t clean up after itself. You’re constantly deleting things it build during debugging. And if you don’t keep a periodically refreshed context file, forget about it. It’s a trade off early on but by the time you scale, all those kinks will be worked out as a partnership between your engineers and your customers.
The sample script I was aiming for is below. After experimenting with it, it was too restrictive. Also I hadn’t done enough pre-processing of the data, like removing words that don’t add much value to a sentence to improve model responses more efficiently. The data pipeline at that time was also missing recommendations on what weight to assign to article quality.
The result was a set of summaries, spoken with conviction, but lacking citation in the way it would be presented in a broadcast format. At scale, eventually not validating that trust with listeners up front feels like it’s too critical a risk to ignore.
Playing around with Temperature, Top-K, and Top-P revealed this model (Titan Lite) out-the-box not be good enough at this time for the end result I’m looking for. Up next is experimenting with a couple of other models available in Bedrock’a marketplace and comparing it to the results from fine-tuning the base model.
I’m ok with hitting my token limit. At least it gives me some time to research and build out some of the other end-state features like user authentication, integrating subscription onboarding, adding advanced security controls, and testing pipeline automations, while setting operations alarms, actions and goals.
Let’s see how far one day a week and one day a weekend will get me heading into the new year.
—
Sample Script – Listen to Today’s Episode
This is a test post demonstrating the new enhanced blog structure for better content discovery.
Topics Covered Today
- AI Regulation Updates from the EU
- Cryptocurrency Market Analysis
- Cybersecurity Breach at Major Tech Company
- New Quantum Computing Breakthrough
Episode Summary
This test episode demonstrates the enhanced content structure for better discovery and navigation. We cover major developments in AI regulation, cryptocurrency markets, cybersecurity incidents, and quantum computing advances.
Episode Stats:
- Sources analyzed: 5 RSS feeds
- Generated: 2025-12-14 at 19:30 UTC
- Voice: Danielle (AWS Polly Neural)
- Duration: 18-20 minutes