Weekend Creative Agent Challenge: Akashvani
A tiny AI radio station that never sleeps. Every day at 10 AM IST, it has a Bollywood or Hollywood character perform a monologue about history.
Vision & What It Does
Akashvani ("voice from the sky," historically the name of India's national radio broadcaster) is a fully autonomous daily theatre segment. Every day at 10:00 AM IST, with nobody watching, it:
- Looks up what actually happened on today's date using Wikipedia's "on this day" feed, preferring events with an Indian connection when one exists.
- Picks one character from a rotating cast of 20 Bollywood and Hollywood icons (Munna Bhai, Gabbar Singh, Basanti, Rancho, and Baburao alongside Tony Stark, Yoda, Gandalf, Sherlock Holmes, and Forrest Gump), while avoiding whoever was used in the last five days.
- Writes a 150-250 word theatre monologue in that character's voice, tying their persona to the day's occasion with wit, irony, or solemnity as the day calls for, complete with stage directions and a radio-style sign-off.
- Stores it and serves it to a small live website, with zero human involvement in the loop.
The problem it solves is exactly the one the challenge names: the best creative tool is the one you never have to open. I don't write a prompt each morning or pick a character; I just check the site, and there's a new, genuinely different piece of micro-theatre waiting, anchored to something real that happened on that date in history. A Forrest Gump monologue about a 1943 Bengal famine reporter, delivered in Gump's folksy cadence with "stupid is as stupid does," is not something I'd have thought to prompt for; it's what the system produced on its own.
How You Built It
I built the whole stack (infrastructure, two Lambdas, and a frontend) using AWS SAM, and deliberately verified everything against a real AWS account rather than trusting the plan on paper.
The generation logic originally called Amazon Bedrock's Nova Pro model through the
apac.amazon.nova-pro-v1:0 cross-region inference profile (the direct on-demand model ID isn't invocable in ap-south-1, confirmed by an actual ValidationException before writing a line of Lambda code). It worked, but a real test invoke hit ThrottlingException: Too many tokens per day ; aws service-quotas confirmed this account's brand-new Bedrock usage had a literal 0.0 daily token quota, an account-level ramp-up limit, not a bug. Rather than wait it out, I swapped the integration entirely for a direct call to the Gemini API.That swap surfaced two more real bugs, both diagnosed the same way: reproduce against the live API directly. First,
gemini-2.5-flash returned a 404: Google had retired it for new keys mid-build, with the error message itself naming gemini-3.6-flash as the replacement. Second, the very first real story came back as one truncated sentence. A raw curl test against the Gemini endpoint showed why: finishReason: MAX_TOKENS with 574 of a 600-token budget silently spent on the model's internal "thinking" before it wrote a single visible word. Setting thinkingConfig: { thinkingLevel: 'low' } and raising the budget fixed it, confirmed with a second direct API call before touching the Lambda again.AWS Services Used / Architecture Overview
Everything is defined in one AWS SAM template and deployed to
ap-south-1 .
Text version of the same flow:
EventBridge Scheduler (cron(0 10 * * ? *), tz=Asia/Kolkata)
│ autonomous daily trigger, no human involved
▼
Generate Lambda (Node.js)
├─ Wikipedia "on this day" API → today's occasion
├─ DynamoDB Scan (last 5 dates) → avoid repeat characters
├─ 20-character roster (static) → random pick, excluding recent
└─ Gemini API (generateContent) → 150-250 word monologue
│
▼
DynamoDB (akashvani-stories): single source of truth
│
▼
Read Lambda -> API Gateway (HTTP API, CORS-enabled) -> S3 static website (Vite + React frontend)Services: Lambda (generation + read), DynamoDB (on-demand billing, one table), EventBridge Scheduler (the actual autonomy: a first-class
AWS::Scheduler::Schedule resource with a native IANA timezone, not a UTC-shifted cron guess), API Gateway (HTTP API), S3 (static site hosting, public read via bucket policy), and IAM (a separate least-privilege role per Lambda: the read function can't write, and neither function has any permission it doesn't use). Google's Gemini API is called directly over HTTPS from the generation Lambda using Node's native fetch , with the API key handled as a NoEcho CloudFormation parameter sourced from a local, git-ignored file; it's never logged or committed.Example
The live site showing a generated story:

CloudWatch log group showing an autonomous run (occasion lookup, character selection, and a successful Gemini call from a scheduled/manual run):

What You Learned
I learned that Bedrock's cross-region inference profiles are a distinct resource from the underlying foundation model, with their own ARN and IAM permission shape, and that a brand-new account's Bedrock quota can start at literally zero tokens per day, which looks identical to a permissions bug until you check
service-quotas directly. I learned that newer Gemini models budget "thinking" tokens out of the same pool as visible output, so a token-truncated response can look like a formatting bug when it's actually a budget problem. And the broader lesson: when output looks corrupted, check it at the source (the database, the API, the raw bytes) before trusting the display layer that showed you the symptom. Link to App or Repo
Live, deployed application (no login required): http://akashvani-atharva-bedekar.s3-website.ap-south-1.amazonaws.com
The generation Lambda runs unattended every day at 10:00 AM IST via EventBridge Scheduler; the site above always reflects the most recent autonomous run.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article