
Weekend Creative Challenge: AI Transformation Playbook
Pitch an AI initiative to four executives who object out loud, in character. Pick a market and an industry, or paste a document and let the app work them out. The same pitch meets a completely different room in Singapore, Tokyo or Jakarta. You leave with a score, the three questions you are least ready for, and the objections read aloud. Built on AWS Lambda, API Gateway, S3, Amazon Bedrock and Amazon Polly.
#creative-expression #challenge #AWS #AmazonBedrock #AmazonPolly
Vision and what the app does
It is a boardroom simulator you can lose.
You type a paragraph describing an AI initiative, and four executives read it and object, out loud, in character. The Chair wants to know who the customer is. The CFO wants the number. The Chief Risk Officer wants to know whose name is on it the day the model is confidently wrong. The CEO thinks it is too small.
Then you get a score out of 100, and the three questions you are least ready for. Most first attempts get sent back before the second question, which is the joke and also the point.
The part I actually care about is what happens before you pitch. You pick a market and an industry, or you paste a document and let the app work them out from the text. The same pitch then faces a different room in each. A Singapore board asks who else in the region already runs this. A Tokyo board asks which departments agreed, and what happens to the people whose work changes. A Jakarta board asks who operates it on Monday once the launch team goes home.
Run the built-in example against Singapore and it scores 100. Run the identical text against Japan and it scores 80. The twenty points it drops are a reference to MAS FEAT, which means nothing to a board in Tokyo.
The board also speaks. Each member has an Amazon Polly voice, and the Chair takes the accent of whichever market you selected: Jasmine for Singapore, Olivia for Australia, Kajal for India. The other three keep fixed voices so you can still tell who is talking. Reading four objections on a page is homework, and I wanted it to feel more like sitting in the room.
How I built it
Most of the design work went into deciding where a model should not go.
The model does not score the pitch. It extracts nine signals, answering questions like whether this text names a customer and where exactly it does so. Deterministic code then decides what those signals are worth. Language understanding is genuinely hard and variable, so it gets a model. The rubric is a rule, and a rule needs to be a guarantee rather than a suggestion, so it gets an if-statement.
Putting the rubric in the prompt would have broken three things. The same pitch would score differently across runs. A confident, well-written pitch with no customer in it would talk its way to a high score, which is the exact failure the app exists to catch. And nobody could audit why a verdict came out the way it did.
Customer obsession is a gate, not only a weight. It carries 40 of the 100 points, the heaviest axis. Weighting alone would still let a strong architecture score buy an approval, so it is also a hard gate: a pitch that does not name a customer and quote what that customer actually said cannot be approved, whatever else it scores. The Chair's opening question is hardcoded and identical in every market, every industry, every run. A board member you can charm into skipping the customer question is not a useful simulator.
There is also no loop anywhere in the system. One model call per request, and the code path makes more than one structurally impossible. That was cheaper than the alternative. A system that cannot loop does not need a loop detector, a step budget, or a kill switch for a runaway agent, which are three things I would otherwise have had to build and then trust.
I treated Bedrock as an enhancement rather than a dependency, because my AWS account was a few hours old when I started. Every request returns a complete hearing whether or not the model answers, and every response carries a
source field that the interface displays. Degrading quietly would have been worse than failing loudly, since you would attribute template output to a model and calibrate your trust wrongly.That decision mattered more than I expected it to.
AWS services and architecture
There are two independent paths on the account. One serves the app on every
request. One runs weekly and maintains reference data.
Request path
Browser (Amazon S3 static website) | POST { pitch, market?, industry? } vAmazon API Gateway (HTTP API, CORS) vAWS Lambda (Node 22) 1 validate deterministic 2 route country x industry deterministic 3 convene the board Amazon Bedrock, falls back to templates 4 resolve context deterministic 5 score + Working Backwards gate deterministic 6 narrate objections Amazon Polly, cached to S3 7 prep sheet + brief deterministicAWS Lambda runs the handler. Amazon API Gateway exposes it. Amazon S3 hosts the three static frontend files and doubles as the audio cache. AWS IAM scopes the function to Bedrock invoke on Nova ARNs, Polly synthesis, and S3 writes limited to the audio prefix of a single bucket. Amazon Bedrock is wired in for the objections. Amazon Polly narrates them.
Polly audio is content addressed. The object key is a hash of the text plus the voice, so identical text in the same voice is synthesised once and every later request becomes a HeadObject rather than a synthesis.
Weekly path
The reference cases in this app are hand curated, which means they decay. So there is a second Lambda, on an Amazon EventBridge schedule, that proposes new ones.
Amazon EventBridge (Mondays) vAWS Lambda (research agent) | reads a GitHub token from AWS Systems Manager Parameter Store vAmazon Bedrock proposes candidate cases vverification gate deterministic, in code - source host must belong to the vendor - URL must return 200 - the page must actually name the organisation - every figure in the outcome must appear in the source vGitHub pull request for a human to reviewThe agent never merges and never pushes to a default branch. Everything the model returns is treated as an unverified claim, and each claim is checked against the live page before it can reach a pull request. If an organisation and a source check out but a claimed figure cannot be found, the case is kept and the figure is dropped, because a real case with no published number is exactly what a cited library should contain.
One honest note on scope. This agent maintains the wider APJC playbook library published separately at rayhan.ai/apjc-playbook. The app carries its own smaller reference set. Same account, same pattern, separate data, and the diagram says so rather than implying a connection that does not exist.
What I learned
Lambda Function URLs returned 403 to the public internet and I never worked out why. Auth type NONE, a textbook
Principal: "*" resource policy, no Organization, no account level Block Public Access, and still AccessDeniedException. A direct aws lambda invoke worked, so the function was fine and only the public edge was refusing. I gave it about forty minutes and then switched to an API Gateway HTTP API, which is public by default. The handler needed no changes at all, because HTTP API sends the same v2 payload shape that Function URLs do.A new AWS account can have zero Bedrock quota, and this is not an access problem. The model access page has been retired and serverless foundation models now enable automatically on first invocation, so my call reached Bedrock and came back with
ThrottlingException: Too many tokens per day . A two token "hi" prompt gets the same error. Service Quotas shows why: every on demand value for Nova Lite reads 0.0, and the ones that matter are marked not adjustable. That is a new account activation state. Two other builders in this challenge's comments are stuck on the same thing.So the app currently serves its template board, badged as such in the interface. I wrote that fallback on day one for a failure I had only imagined, and hit the real version of it the same afternoon.
Amazon Polly is the reason this is still a creative app. With Bedrock sitting at zero I went looking at the other AI services. Comprehend and Translate both refused with SubscriptionRequiredException. Polly answered on the first call. It also has a Singapore English voice, so the Singapore board now sounds Singaporean, which does more for the premise of the app than anything I had planned.
The bug I am most glad I found had nothing to do with AWS. My frontend defined a local
escape() helper for HTML escaping, which silently shadowed the global escape that my permalink encoder was calling. Every shareable link produced garbage, failed JSON.parse , got swallowed by a catch block, and did nothing at all. It only surfaced when I bundled the app into a single file and started exercising paths that localhost never had. The old global string functions are easy to shadow without noticing, and nothing warns you. TextEncoder and a helper renamed to esc fixed it.Links
Pick a market and an industry, load one of the example pitches, and press the play buttons.
Built on AWS Lambda, Amazon API Gateway, Amazon S3, AWS IAM, Amazon Bedrock and Amazon Polly.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article