How I Built a Full-Stack AI App on AWS in 4 Days as a Complete AWS Beginner
A final-year B.Tech CSE student's journey of building and deploying a full-stack AI-powered kitchen management app on AWS in four days — from zero AWS experience to six Lambda functions, DynamoDB, Amazon Bedrock Nova Micro, SNS, and EventBridge.
Wait, What Did I Just Do?
Four days ago, I had never touched AWS in my life. I did not know what Lambda was. I had never deployed anything beyond a college assignment. I did not know what API Gateway meant or why anyone would need it.
Today, I have a live, deployed, full-stack application running on AWS with six Lambda functions, a DynamoDB database, an AI model from Amazon Bedrock, automated email alerts via SNS, and a scheduled job running every night at 2 AM via EventBridge.
I am a final-year B.Tech CSE student from Kolkata. And I built all of this in four days, solo, for the First Commit hackathon by WeMakeDevs and AWS.
This is the story of how it happened.
The Idea — Where It Came From
The idea did not come from a research paper or a trend report. It came from my own kitchen.
In every Indian household I know, someone mentally keeps track of everything in the fridge. They know what was bought, roughly when, and what needs to be used first. But sometimes things slip — a packet of paneer hidden behind other things, leftover dal from two days ago, milk that was supposed to be used yesterday. And then the food goes bad and gets thrown away.
And it is not carelessness. It is just that there is no system. You cannot remember everything.
When I saw the hackathon theme — build something that solves a real problem — I thought of this immediately. India wastes 78 million tonnes of food every year. A huge part of that happens at the household level, not at the supply chain level. And nobody had built a simple, India-specific app to fix it.
That became KitchenSense.
Day 1 — 17th September | The Backend
I started with Java because that is what I know best. I am targeting Software Engineering roles after graduation and Java Backend development is my specialization, so this felt like the right choice.
The architecture I had in mind was serverless — AWS Lambda functions behind an API Gateway, connected to DynamoDB. I had read about this pattern but never actually built it.
On Day 1, I wrote all six Lambda handlers:
- AddItemHandler — saves a new food item to DynamoDB
- GetItemsHandler — fetches all items for a user, calculates days until expiry
- DeleteItemHandler — removes an item when marked as consumed
- RecipeHandler — calls Bedrock to suggest a recipe from expiring items
- ChatHandler — handles AI chat with the user's inventory as context
- NotificationHandler — sends daily SNS alerts for expiring items
I also wrote the FoodItem model, DynamoDBService, BedrockService, and SNSService.
The most interesting part of the backend was the
getEffectiveExpiry() method in FoodItem. For packaged items, the expiry date comes straight from the packet. For homemade food like dal or sabzi, there is no printed date — so I compute it from the cook date plus shelf life days. This felt like a genuinely useful design decision for Indian kitchens where you cook fresh food regularly.The hardest part of Day 1 was understanding how DynamoDB works. Coming from a relational database background, the concept of partition keys and sort keys took me a while to understand. But once it clicked — userId as partition key, itemId as sort key — everything made sense. It is elegant actually.
Day 2 — 18th September | The Frontend and AWS Console Setup
I am not primarily a frontend developer. But I wanted KitchenSense to look good — not like a college project, but like something people would actually want to use.
I built six pages:
- Landing page — hero section with the value proposition
- Auth page — login and signup with a fun character that looks away when you type your password
- Dashboard — the main kitchen view with stats, item list, AI recipe card
- Add Item page — form for packaged and homemade items
- Chat page — AI chat with a context panel showing urgent items
- Profile page — kitchen settings and notification preferences
A note on authentication — I chose not to implement AWS Cognito for this hackathon. Cognito is the right production choice for user authentication, and I did set it up initially. But integrating Cognito into a frontend without a framework, managing JWT tokens, handling refresh flows, and wiring it to all six Lambda functions would have taken the entire weekend on its own. For a four-day hackathon, participating as a solo developer, I made a deliberate decision to simulate authentication using localStorage and focus my time on the features that actually demonstrate the value of KitchenSense — the AI integration, the expiry tracking, and the real-time dashboard. In a production version, Cognito would replace the localStorage auth completely, and the userId would come from the validated JWT token instead of being generated from the email address.
The character on the auth page that looks away when you focus on the password field — that was a small detail I added at 2 AM and I am genuinely proud of it. It made the app feel alive.
While building the frontend, I also started setting up AWS:
- Created the DynamoDB table with the right keys
- Created the IAM role for Lambda with permissions for DynamoDB, Bedrock, SNS, and CloudWatch
- Created the SNS topic for email alerts
- Created all six Lambda functions with Java 21 runtime
- Created the API Gateway with four resources — /items, /items/{id}, /chat, /recipe
- Set up EventBridge to trigger the notification Lambda every night
Setting up API Gateway was the most confusing part. There are so many settings — methods, integrations, CORS, stages, deployments. I made mistakes and had to redo things multiple times. But by the end of Day 2 I had a working API structure.
Day 3 — 19th September | Testing, Debugging, and Everything Breaking
Day 3 was the most stressful and also the most educational day.
I hosted the frontend on S3 as a static website and opened it in the browser. It looked beautiful. And then nothing worked.
Here is a list of everything that broke and how I fixed it:
Problem 1 — CORS errors
The browser was blocking all API calls because API Gateway was not sending the right CORS headers. I had to go into every single resource in API Gateway, enable CORS, and redeploy. Multiple times.
Problem 2 — Lambda Proxy Integration was off
The API was returning a weird response with statusCode, headers, and body as separate fields instead of a proper HTTP response. This was because Lambda Proxy Integration was not enabled. Once I turned it on for every method and redeployed, everything started working.
Problem 3 — /recipe was a POST but the frontend called GET
I had set up /recipe as POST in API Gateway but my JavaScript was making a GET request. Fixed by deleting the POST method and creating a GET method instead.
Problem 4 — Bedrock model not available
I had built the backend using Amazon Titan Text Express. When I tested it, I got errors. I went to the Bedrock console and found that the Model Access page had been retired — Titan was no longer available in the old console. I had to switch to Amazon Nova Micro using the new Converse API. This required rewriting the invokeModel method and updating the AWS SDK version from 2.25.11 to 2.26.12.
Problem 5 — fridgeStatus values mismatched
The frontend was storing
standard , adjusted , pantry in localStorage. The backend Bedrock prompt was expecting FUNCTIONAL , OLD , NONE . The AI was getting confused. I fixed this by adding a mapping function in api.js.Problem 6 — add-Item.html had a capital I
S3 is case-sensitive. Windows is not. My file was saved as
add-Item.html on Windows but the links in my HTML said add-item.html . On S3, these are two different files. The app was redirecting to the landing page every time someone clicked Add Item. Fixed by renaming the file and reuploading.By evening, everything was working. The dashboard was loading items from DynamoDB. The AI was generating real Indian recipes. The chat was responding intelligently based on actual inventory. Mark as Consumed was removing items instantly.
I tested the full flow end to end and it worked. I recorded the demo video.
Day 4 — 20th September | The Blog, The Video, The Submission
Today is the final day. I edited the demo video in Microsoft Clipchamp, added voiceover, and uploaded it to YouTube. I wrote this blog. I am about to submit.
It is a lot to do in one day. But I am doing it.
What I Actually Learned
I came into this hackathon knowing Java but knowing nothing about AWS. Here is what I know now:
Serverless architecture — I understand what it means to build without managing servers. Lambda functions scale automatically, you pay only for what you use, and the cold start problem is real (my Lambda functions take 3-4 seconds on first invocation with Java 21).
DynamoDB — NoSQL design is different from relational design. You design your table around your access patterns, not around normalisation. Choosing userId as partition key and itemId as sort key was the right call for my use case.
API Gateway — CORS is not optional. Lambda Proxy Integration is not optional. Deploying after every change is not optional. I learned all of these the hard way.
Amazon Bedrock — The Converse API is clean and simple. The prompt engineering matters more than the model settings. A well-written prompt that gives the AI context about Indian kitchens and the user's fridge status produces much better recipes than a generic prompt.
AWS SDK v2 — The builder pattern used throughout the SDK takes getting used to. But once you understand it, it is consistent and readable.
Debugging a live deployed application — This is very different from debugging locally. CloudWatch logs are your only window into what is happening. Learning to read them quickly was essential.
What I Am Taking With Me
Beyond the technical skills, I am taking something else with me from this weekend.
I proved to myself that I can build something real. Not a tutorial project. Not a clone. Something original, deployed, working, solving a real problem.
I am a final-year student targeting Software Engineering roles. Before this weekend, my portfolio had college projects. Now it has a live AWS deployment with a real AI integration.
That matters.
The Stack, One More Time
For anyone who wants to build something similar:
- Frontend — Pure HTML, CSS, JavaScript hosted on AWS S3
- Backend — Java 21 on AWS Lambda
- Database — Amazon DynamoDB
- AI — Amazon Bedrock Nova Micro via Converse API
- Notifications — Amazon SNS + Amazon EventBridge
- API — Amazon API Gateway (REST)
- Build — Maven, AWS SDK v2 (2.26.12)
Total AWS cost for the weekend — close to zero on free tier.
AI Tools That Helped Me Build This
This hackathon allowed the use of AI coding tools, and I used several of them. Here is honestly what each one contributed:
Claude (Anthropic) was my primary technical partner throughout the four days. I used it for the overall system architecture, reviewing my Java Lambda handlers for correctness, guiding me through every AWS console setup step, debugging the CORS issues and Lambda Proxy Integration problems, fixing the Bedrock model switch from Titan to Nova Micro, and reviewing both the frontend and backend code end to end. When things broke at 10 PM on Day 3, Claude walked me through every fix one by one.
Gemini (Google) helped with code generation for parts of the backend service layer and some of the initial Java boilerplate.
ChatGPT (OpenAI) helped with frontend code fixes — particularly around the responsive CSS, the unified design system across all five pages, and some of the JavaScript logic in the dashboard and chat pages.
Google Stitch was used for the initial UI/UX design and mockups — the visual language, colour palette, and layout structure of KitchenSense came from Stitch before I translated them into actual HTML and CSS.
I want to be honest about this because the hackathon rules require it, but also because I think it is worth saying: AI tools did not build KitchenSense. They helped me build it faster and catch mistakes I would have missed. Every architectural decision, every design choice, every debugging session — those were mine. I had to understand every piece of code to integrate it, fix it when it broke, and explain it in this blog. That understanding is real, even when the first draft came from an AI.
Try It
The full source code is available on my GitHub:
Feel free to explore the code, the architecture, and the AWS setup. If you have questions, reach out on LinkedIn.
Final Thought
If you are a student who has never used AWS and you are wondering whether to enter a hackathon like this — do it. You will break things. Everything will stop working at the worst possible moment. You will fix it. And on the other side of that, you will have built something real.
That is the whole point.
Srija Ghosh is a B.Tech CSE student at Supreme Knowledge Foundation Group of Institutions (MAKAUT), Kolkata. She is targeting Software Engineering roles and can be found on LinkedIn and GitHub .
Built for First Commit — Bharat Builds Tour by WeMakeDevs × AWS, September 2026.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article