AWS Builder Center
Weekend Creative Challenge: Two Nouns, One Unicorn

Weekend Creative Challenge: Two Nouns, One Unicorn

A web app that generates satirical startup pitches. You get a fictional tech company (name, tagline, pitch, buzzwords, funding stage, valuation, and a pretentious founder quote) rendered as a polished pitch-deck card. The startups were pre-generated with Amazon Nova Pro and are served as a static site on Amazon S3.

Cloud Engineer

Vision and what the app does

Every startup pitch sounds the same after a while. Same buzzwords, same hockey-stick charts, same founder staring into the middle distance. I wanted to make fun of that, and I wanted the joke to be reusable, so I built a machine that generates a fake startup from any two random words.
You type two nouns. Otter and stapler. Glacier and banjo. You can also hit "Surprise me" and let the app pick for you. A second later you get a full company: a brandable name, a punchy tagline, a made-up industry category, a straight-faced elevator pitch, four trendy buzzwords, a funding stage, a playful valuation, a vanity metric, and a pretentious founder quote. It all lands on a designed pitch-deck card with a generated monogram logo.
The output is the creative part. "GlacierBanjo, Cryo-String Fusion Solutions, melodic climate resilience, Series A, valued at $34.5M with a snowflake on it." The model commits to the bit completely, and that deadpan confidence is what makes people laugh and immediately generate another one.
Feed it "noodle" and "glacier" and you get NoodleIce, a Cryo-Culinary Innovation company "revolutionizing sustainability with frozen pasta solutions," complete with a 98.7% Glacial Noodle Retention Rate.
The generated startup card for NoodleIce, showing the logo, tagline, pitch, buzzword chips, funding stats, and founder quote


How I built it

I started from the model, not the UI, because the whole app is only as good as the text it generates.
The app's input screen

I wrote a system prompt that casts the model as a deadpan venture-capital satirist and forces it to return one strict JSON object with fixed keys. Locking the shape of the output early meant the frontend never has to guess what it received.
The backend is a single Python function. It validates the two nouns, caps their length, calls the model through the Converse API with a high temperature so results stay varied, then parses the JSON and hands it back. I added a small fallback that extracts the JSON block if the model ever wraps it in stray text, plus adaptive retries for throttling.
The frontend is one HTML file. No framework, no build step. It renders the card, derives the logo color from the company name so each result looks distinct, and escapes every field before injecting it into the page.
The challenges were real. My first plan used an image model for logos, but the image generation model on my account had been retired to legacy status and was locked, so I pivoted to generating the logo as a CSS monogram instead. That turned out better and faster. The second problem was a CORS bug: my function set the CORS headers and the Function URL also set them, so the browser saw two values merged into one invalid header and blocked every request. The fix was to set CORS in exactly one place and let the Function URL own it.

AWS services and architecture

  • Amazon Bedrock (Nova Pro) generates the startup as structured text through the Converse API.
  • AWS Lambda runs the generator on a public Function URL, so there is no separate API layer to manage.
  • Amazon S3 hosts the static site.
  • IAM gives the function a role scoped to invoking one model and nothing else.
Browser (S3 static site)
| HTTPS POST { noun_a, noun_b }
v
Lambda Function URL -> Lambda (Python)
| Converse API
v
Amazon Bedrock: Nova Pro
|
v
strict JSON -> rendered card
The request path is short on purpose. One hop from the browser to Lambda, one hop from Lambda to the model. Fewer moving parts meant I could deploy in an afternoon and debug the two issues quickly.

What I learned

Forcing a model to return strict JSON is the difference between a demo that works and one that breaks on stage. The prompt does more heavy lifting than the code.
I also learned to check model access before committing to it. A model can appear in the catalog and still be locked because it was marked legacy and my account had not used it recently. Finding that out early saved the project, and dropping image generation for CSS made the app both cheaper and more polished.
Lambda Function URLs are the fastest way to put a model behind a browser. Just remember that if you configure CORS on the URL, do not also set those headers in your code.
The app's input screen


Link

Try otter and stapler first. Then let it surprise you.
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