AWS Builder Center
Weekend Showcase Challenge: Verse

Weekend Showcase Challenge: Verse

A daily weather-poem agent that now survives Bedrock being unreachable — falls back to a local poet instead of failing silently.

Weekend Showcase Challenge: Verse

Tagging  Lewis Sawe  as the builder who inspired me this summer - his "Museum That Grows" (an agent that accumulates a fictional civilization's artifacts one memory-aware day at a time) showed me a version of "always-on" I hadn't considered: an agent whose output compounds over time instead of resetting each run.

Vision and what it does

For the always-on agent challenge, I built Verse -a daily weather-themed poem generator. Every morning, before anyone opens a laptop, it checks the sky over Bengaluru, picks a poetic form for the week (haiku, sonnet, free verse, villanelle, ode - rotating so the "voice" visibly shifts over the run instead of staying static), and writes an original poem capturing that day's weather. It lands in a small gallery site styled like a weather almanac: one entry per day, stamped with date, temperature, and form.
That version shipped, but it had a gap I wasn't proud of: it depended entirely on Amazon Bedrock being reachable, and mid-challenge I hit an account-level Bedrock quota that AWS Support still hasn't resolved - every model on my account, not just the one I'd chosen, was returning  ValidationException: Operation not allowed . The agent was correctly built and fully deployed, but it couldn't actually write anything.
For this showcase weekend, I used the "add the finishing touch you ran out of time for" option and fixed exactly that gap - not by working around Bedrock, but by making the agent resilient to Bedrock being unavailable at all.

How I built it

The fix is a single, honest design decision: the generator Lambda now tries Bedrock first, and if that call fails for any reason - a quota block like mine, a transient outage, a region issue - it falls back to a local, deterministic poem generator instead of failing the run. The fallback builds lines from weather-mapped word banks (separate banks for clear skies, rain, storm, fog, snow, cloud) using templates, seeded by the date so a given day always produces the same poem even across retries. Every stored poem now carries a  source  field -  "bedrock"  or  "fallback"  - and the frontend displays which one produced it, right on the card. Nothing is hidden.
The harder part wasn't the code, it was diagnosing what was actually wrong in the first place.  ValidationException: Operation not allowed  gives you almost nothing to go on. I worked through it methodically: confirmed IAM permissions were correct, confirmed every Bedrock model (not just Nova Lite) failed the same way, which ruled out anything model-specific, then found the real answer in the Service Quotas console - the on-demand inference quota for Nova Lite was explicitly set to 0 and marked "Not adjustable," meaning only AWS Support could raise it. That's now an open case. Rather than let a support ticket's timeline decide whether my project works, I decided the agent itself should be the thing that doesn't fail -regardless of what AWS's side is doing.

AWS services used and architecture

  • EventBridge Scheduler -fires the generator daily, no human involved
  • AWS Lambda (Node.js) - generator  (weather → Bedrock-or-fallback → DynamoDB) and  reader  (DynamoDB → API)
  • Amazon Bedrock (Nova Lite) - primary poem source, used whenever reachable
  • Amazon DynamoDB - one item per day: poem, weather, style, and generation source
  • API Gateway (HTTP API) - exposes  GET /poems 
  • Amazon S3 -hosts the static gallery frontend
Flow:  EventBridge → Generator Lambda → (Bedrock, with local fallback) → DynamoDB , separately  S3 frontend → API Gateway → Reader Lambda → DynamoDB . The write and read paths never share a Lambda, which kept each function's IAM policy narrow.

What I learned across the summer

This summer's builds -from an Odoo hackathon transport-ops platform, to a GitLab Duo agent pipeline reviewing merge requests, to Verse -all taught the same underlying lesson from different angles: the interesting engineering problem is rarely the happy path. TransitOps needed row-level locking because concurrent dispatch is where real systems break. My GitLab agent work needed careful staged pipelines because a single agent call isn't trustworthy enough on its own. And Verse, this weekend, needed to accept that the AI model it was built around might just not be there - and still do its job anyway.
If there's one thing I'd tell myself back in august, it's that "does it work" and "does it keep working when something outside your code breaks" are different questions, and the second one is usually the more valuable thing to have actually solved.

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