AWS Builder Center

Weekend Creative Agent Challenge: Terra Pulse

An always-on agent that writes the Earth's daily diary. Every morning it reads the last 24h of global earthquakes from USGS and Amazon Nova Micro on Bedrock ghostwrites a first-person chronicle as the planet itself, published on a public page before you wake up. Built for anyone who lives over a fault line and checks seismic feeds more than they should.

Data Engineer

Why I built this

I keep a browser tab open on the USGS earthquake map more often than I'd like to admit. If you live anywhere near the Pacific Ring of Fire you probably do too. This year made that habit hard to drop: in June, a magnitude 7.8 offshore quake hit Mindanao in the Philippines and sent a meter-high tsunami into coastal villages, on what happened to be the first day of school. Seven weeks ago, Flores in Indonesia took a 7.7 that killed dozens and left tens of thousands sleeping outside because of aftershocks. And earlier this year Venezuela had the deadliest earthquake of 2026 so far, with over six thousand lives lost.
When you follow this stuff regularly, you notice something: raw seismic lists are accurate and completely joyless. A table of magnitudes and epicenters tells you nothing about how to feel, and after big events people refresh those feeds at 3 a.m. looking for something the data can't give them: context, rhythm, reassurance that most days the planet is just quietly stretching.
So for the challenge I built Terra Pulse, an agent that writes the Earth's daily diary. Every morning at 10:00 UTC an EventBridge rule wakes up a Lambda function with no human involved. It pulls every earthquake of magnitude 4 or higher from the previous 24 hours through the USGS public API, sends the list to Amazon Nova Micro on Bedrock, and asks it to write a short journal entry in first person, as the planet itself. The entry gets stored in DynamoDB together with the structured quake data and shows up automatically on a public page. You never open anything. You wake up and yesterday's entry is already written, signed "~ Earth".

Who it's for

Two groups, honestly. People who live on top of fault lines and deal with that background anxiety, where every unusual rumble means reaching for the phone. For them, reading a calm narrative every morning builds a sense of the planet's normal baseline, which makes real anomalies easier to recognize instead of just scary. And people far from home who still have family in seismic regions, for whom a one-minute read about the ground under their parents' house is a small ritual worth having.
One thing I want to be explicit about because it matters: Terra Pulse is not an early warning system and is not trying to be. Official alerts from USGS, PHIVOLCS, BMKG and every national agency are the authoritative source and they save lives. This sits next to them, doing the thing data alone doesn't, which is helping people pay attention without burning out.

How I built it

I set two constraints for the weekend. Everything had to fit inside the Free Tier, and the whole stack had to come up from one script,  deploy.sh , written in plain AWS CLI calls so anyone could reproduce it during lunch. No CloudFormation, no CDK, no console clicking.
A few decisions worth mentioning:
  • One Lambda plays both roles. EventBridge invokes it to generate the entry, and the same function serves the public page when someone hits its Function URL. JSON lives at  ?format=json . Two functions would've been two things to babysit for no benefit at this size.
  • Nova Micro runs through a cross-region inference profile ( us.amazon.nova-micro-v1:0 ) instead of a bare model ID, which sidesteps the direct-invocation deprecation.
  • Zero dependencies. The USGS call uses Python's stdlib urllib, so the deployment package is literally one file.
  • Quiet days are handled explicitly. Some 24-hour windows have zero quakes above magnitude 4, so the prompt tells the model to reflect on the silence instead of crashing on empty input.
The hardest part was tone, not infrastructure. My first prompts produced either dry bullet points or geological melodrama. What fixed it was giving the model a strict format contract: Title, Body, Signature ending with "~ Earth". Structure turned out to enable creativity rather than limit it, which was not what I expected going in.

Architecture

Terra Pulse Architecture

EventBridge Scheduler for the heartbeat, Lambda for generation and serving, Bedrock with Nova Micro as the ghostwriter, DynamoDB for persistence, IAM scoped down to one table plus invoke permissions. Everything is event-driven, nothing runs unless something happens, and the projected monthly bill rounds to zero.

What I learned

Agents earn trust by being honest about what they're not. Writing "this will never replace official alerts" into the project made the actual value proposition clearer, not weaker. I also learned that prompting an unattended LLM is a different sport from chatting with one: there is nobody watching to regenerate a bad output, so format contracts and empty-input handling stop being optional details. And finally, that serverless pieces click together surprisingly fast once everything is event-driven, even for someone whose day job is mostly pipelines and tables.
Mostly though, I learned the planet writes a better daily column than I expected.

Try it / Source

Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article