
Weekend Creative Agent Challenge: Postcards from Bengaluru
An always-on AWS agent that reads Bengaluru’s real weather each morning, paints an original postcard, critiques its own artwork with Amazon Nova Lite vision, and redraws until it has something ready in your inbox and gallery before you open a thing.
Vision
The best creative tool is the one you never have to open. Most “daily art” bots are a cron job wrapped around a prompt: they fire, dump text into a bucket, and call it a day. That is automation. It is not an agent.
Postcards from Bengaluru is an always-on creative agent that wakes itself every morning at 06:30 IST, reads the city’s real weather, invents a visual style it has not used in a week, paints an original postcard, then looks at its own artwork with a vision model and redraws once if the work falls short of its own rubric. When I open my inbox, the postcard is already waiting — title, poem, weather context, curator score, and a link to the live gallery.
The problem it solves is simple to state and hard to ship: how do you keep generative work alive without human babysitting? An agent needs three things a script does not have — perception of the real world, memory of what it already made, and a feedback loop that can change tomorrow’s behaviour based on today’s failure. This project is built around those three loops.
Over a week of runs you can watch the gallery evolve. Styles do not repeat. Scores move. Rejected drafts sit next to the winners. The archive is the proof that the agent is improving its judgement in public, not just regenerating noise.
What it does
Every morning the agent:
- Grounds itself in Bengaluru’s live conditions via Open-Meteo — temperature, humidity, cloud cover, wind, rainfall, sunrise and sunset — plus a locally computed moon phase.
- Recalls the last seven postcards from DynamoDB and bans those visual styles so it cannot fall into a rut.
- Art directs a structured scene with Amazon Nova Lite: title, four-line poem, palette, sky gradient, ridges, motif, rain, haze, grain.
- Paints by asking Bedrock for an image first (Nova Canvas / Stable Image dialects). When Bedrock refuses — Legacy models, Marketplace entitlement, missing payment instrument — it falls back to a dependency-free Python rasteriser and draws the postcard itself.
- Critiques the finished PNG with Nova Lite’s vision capability against mood match, composition, depth, and colour harmony.
- Redraws once if the verdict is fail, folding the curator’s revision hint into a revised scene.
- Publishes to a public S3 gallery and emails an HTML postcard with the artwork embedded.
Nobody opens a dashboard. Nobody clicks Generate. The work is simply there when you return.
Architecture

AWS services
Service | Role in the agent |
|---|---|
Amazon EventBridge Scheduler | Fires the agent daily at cron(30 6 * * ? *) in Asia/Kolkata |
AWS Lambda (Python 3.13) | Runs the full agent loop with no third-party packaging |
Amazon Bedrock — Nova Lite | Art direction via Converse + tool schemas; vision critique of the finished PNG |
Amazon Bedrock — image models | Preferred paint path; graceful fallback when the account cannot invoke them |
Amazon DynamoDB | Style memory and postcard archive |
Amazon S3 | Image store + public static website gallery |
Amazon SES | HTML morning email with the artwork embedded |
AWS CloudFormation | One template deploys the whole stack in us-west-2 |
Region choice was not arbitrary. Nova Canvas is marked Legacy and denied to accounts that have not used it recently;
us-east-1 currently exposes no usable text-to-image model for this account; us-west-2 is where Nova Lite is available on demand with image input, so the same model can art direct and see. Everything stays in one region.
How I built it
Process
I started from the challenge brief — an always-on agent that makes something new and has it ready — and rejected the easy path of “Lambda + one Bedrock text call + dump JSON.” Level 200 means perception, memory, and self-critique, not a scheduled prompt.
The build sequence was deliberate:
- Scaffold CloudFormation for S3, DynamoDB, Lambda, Scheduler, and notifications.
- Wire weather grounding and DynamoDB style memory before any generative call.
- Force schema-valid output with Bedrock Converse
toolConfiginstead of hoping for parseable prose.
- Add the vision critic and a hard one-repaint ceiling so cost and runtime stay bounded.
- Survive the image-model wall by implementing a real Bedrock paint path and a local rasteriser fallback.
- Replace SNS email with SES after Gmail’s link scanner kept auto-unsubscribing the SNS subscription.
Key decisions
Structured tools over free-form JSON. Both art direction and critique use Converse with a tool schema and
toolChoice: any . The model returns a toolUse block shaped like the postcard, not a blob of markdown with braces. Sub-scores are re-added in code; if the model’s total disagrees with the arithmetic by more than ten points, the arithmetic wins. Models are better at judging than at adding up.Memory as a hard constraint. Banned styles from the last seven days are not a soft suggestion. Six consecutive runs produced a dawn woodcut, a dawn watercolour, a dusky watercolour, an overcast silkscreen, a misty morning woodcut, and a heat dust / twilight mosaic — distinct moods driven by real weather, not a roulette wheel of adjectives.
Renderer that overrules the model. Language models reliably place a sun below the horizon, bury clouds behind hills, and reverse aerial perspective. The local rasteriser enforces physical rules: sun stays in the sky, clouds lift above the horizon, ridges sort by luminance, distant land blends toward the horizon colour. The agent still chooses what to draw; the renderer refuses to draw nonsense.
Honest graceful degradation.
painter.py really calls InvokeModel with Nova Canvas and Stability request dialects. When Bedrock returns ValidationException or Marketplace entitlement errors, CloudWatch logs the refusal and the agent draws locally instead of skipping the day. Every postcard records its render_source . That is resilience, not a fake call path.SES instead of SNS. SNS’s injected one-click unsubscribe URL was being prefetched by Gmail’s link scanner, which silently deactivated the subscription within minutes. SES delivers a clean HTML card with the artwork embedded — better for the user, and better evidence for this write-up.
Challenges
The hardest challenge was not writing prompts. It was discovering that this account had no usable Bedrock text-to-image model: Nova Canvas Legacy denial, empty image inventory in
us-east-1 , Marketplace INVALID_PAYMENT_INSTRUMENT for Stability. Waiting on a payment method would have burned the weekend. Building a scene rasteriser and PNG encoder in pure Python — scanline fills, soft cloud masks, radial glow, curved palm blades, grain, vignette — kept the agent honest to the brief: always on, always producing, still feeding a real raster into Nova Lite vision.The second challenge was email reliability. SNS looked free and simple until it started deleting itself. Moving to SES after identity verification fixed delivery and upgraded the artefact from plain text to a postcard you can actually open.
Screenshots
1. Live gallery — the agent’s public memory

2. Curator modal — the agent judging itself

3. Morning email — ready before you open anything

Watching improvement over time
A single postcard is a demo. A gallery is an argument.
Across the first week of history the archive already shows:
- Style diversity under constraint — woodcut, watercolour, silkscreen, risograph / mosaic moods without repetition.
- Weather fidelity — monsoon and overcast days pull darker skies and rain streaks; clearer heat days push warm gradients and haze.
- Self-critique that bites — every seeded day ran two attempts; the redraw won on multiple dates. Scores are not decorative 95s; the critic parks competent work in the 50–70 band and says why.
- Public failure — rejected drafts remain visible. Improvement is not “the model got luckier”; it is “the agent revised against a concrete hint and tried again.”
That is what “improves its style over time” looks like when you take the challenge phrase literally: memory changes the next brief, critique changes the next paint, and the gallery accumulates the trail.
What I learned
- An agent is a closed loop, not a schedule. Cron without perception and judgement is just delayed execution. The vision critic and the style ban are what make this Level 200.
- Bedrock Converse tool use is the difference between “usually JSON” and “always a postcard.” Schema enforcement turned the art director and curator into reliable subsystems instead of fragile parsers.
- Account reality beats model marketing. Free Tier credits do not unlock Marketplace models without a payment instrument, and Legacy Nova Canvas will simply refuse you. Designing for refusal — try Bedrock, fall back, log it, keep going — is part of shipping agents on real AWS accounts.
- Operational details kill demos. SNS unsubscribe prefetching, region inventory gaps, and image-model entitlement errors are not footnotes; they are the build. Documenting them is how other builders avoid losing a weekend.
- Cheap can still be ambitious. With the local renderer absorbing paint cost, daily spend is pennies of Nova Lite tokens plus Free Tier Lambda, DynamoDB, S3, Scheduler, and SES. The expensive idea was the architecture, not the bill.
Links
Resource | URL |
|---|---|
Live app | http://postcards-bengaluru-445527450587.s3-website-us-west-2.amazonaws.com |
GitHub repository | https://github.com/harshendram/aws-weekend-challenge |
Challenge | Set Your Creative App Free Weekend Challenge |
Clone, verify your SES identity, deploy with
.\deploy.ps1 -NotifyEmail you@example.com , confirm the schedule in EventBridge, and the next morning’s postcard will be waiting without you opening a thing.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article