A few weeks ago I joined the Decoding Data Science’s 8-Day Building AI Application Challenge. I placed in the top 25 from 295 participants worldwide, what I walked away with is worth more than a podium finish. 8 days, one idea, a fully deployed AI application built from scratch. I built Cookable, an AI powered recipe assistant that takes whatever ingredients you have and generates 5 detailed cookbook quality recipes in seconds. Built with Streamlit, Groq API, and LLaMA 3.3 70B. Here is what I actually learned: • Prompt engineering is harder than it looks. Small wording changes had a bigger impact on output quality than anything else. • Error handling is not optional. 20 test cases on Day 4 found three failure modes I had not anticipated. Each one taught me something. • The gap between working and polished is almost entirely CSS and session state. Core logic was done by Day 3. The rest was making it something I was proud to share a link to. • Deploying something real changes your standards. The moment the app went live, edge cases stopped being optional. I did not get a leaderboard position but I took home something better: a deployed app, a sharper skillset, and a clearer understanding of what it actually takes to build something real from scratch. That feels like enough. Grateful to Decoding Data Science and Mohammad Arshad for building a challenge that forces you to ship, not just study. Live app: https://cookable-bydurgaanand.streamlit.app/ GitHub: github.com/Durga-Anand-2006/cookable Decoding Data Science DDS Business Circle #DecodingDataScience #BuildingAIApplicationChallenge #AIChallenge #BuildInPublic

8 days ago I had a blank Python file and an idea. Today I have a deployed, working AI application that real people can open in a browser right now. ✨ 🥗 Cookable is an AI powered recipe assistant built around a problem most of us have faced. You open your fridge, you have ingredients, you have no idea what to make, and you end up ordering food anyway. 📝 You type in whatever is in your fridge, set your preferences, and the app returns 5 cookbook quality recipes with exact heat levels, timings, and visual cues. Not summaries. Actual instructions you can cook from. Here is what Cookable can do: 🍳 Generate 5 detailed recipes from whatever ingredients you have and pantry staples 🥙 Filter by cuisine, time available, meal type, serving size, and dietary restrictions 🛒 Get 10 smart ingredient suggestions with reasons to buy them 🧈 Generate 2 bonus recipes when you pick up new ingredients 🕘 One click history pills to reload any past search instantly 🌱 Clean earthy green and cream UI built entirely with custom CSS in Streamlit Here is what the 8 days actually looked like: 🫑 Day 1: Ideation. The idea came from standing in front of my fridge at midnight with eggs, a carrot, and half a block of cheese, and ordering food anyway. 🥦 Day 2: Environment setup. Groq API connected, Streamlit running locally, first test prompt returning a response. 🥒 Day 3: The hardest day. Writing prompts that return consistent structured JSON with cookbook quality steps every single time took far more iteration than expected. Small wording changes had a bigger impact on output quality than anything else. 🫛 Day 4: Evaluation and hardening. 20 test cases, three failure modes found and fixed, rate limit handling added, input validation tightened. 🍏 Day 5: Full UI. Custom CSS design system built entirely in Streamlit. Green and cream palette, hero image, history pills, bonus recipe section, sticky footer. 🥬 Day 6: Security. System and user message split, prompt injection guard, retry logic, JSON fallback parser, README cleanup.

Cookable is live. You can open it in any browser, type in whatever is in your fridge, and get 5 detailed recipes in seconds. No login, no setup, just ingredients in and meals out. Deployed on Streamlit Community Cloud, connected directly to the GitHub repo so every future push updates the app automatically. Set the API key securely through the secrets manager, ran the full smoke test checklist on the live URL, and everything passed. Check it out: cookable-bydurgaanand.streamlit.app @decodingdatascience @ddsbusinesscircle #DecodingDataScience #BuildingAIApplicationChallenge #AIChallenge #BuildInPublic

Today was security, debugging, and getting Vault production-ready. 🛠️ What shipped today: ✅ Full security review — API keys in Streamlit Secrets, RLS enforced, no PII in logs ✅ Rebuilt the landing page from scratch ✅ UI polish across Upload, Dashboard, and History ✅ Regenerate roast button on History — with a 30-second cooldown to prevent API spam ✅ Fixed double sidebar bug (Streamlit's default nav conflicting with custom nav) ✅ Fixed button text contrast across all pages ✅ Fixed Pillow version incompatibility blocking deployment on Python 3.14 ✅ 5 real bugs found and fixed — including edge cases like zero income and negative savings ✅ README fully updated — folder structure, schema, security notes, known limitations 🚀 Where things stand: Vault is deployed and live. Every page works end to end on the live URL. One raccoon. Real spending data. No mercy. ⏭️ Tomorrow: final smoke tests, polish, and submission. #AIChallenge #BuildInPublic #Vault #DecodingDataScience #DDSBusinessCircle #BuildAIAppChallenge

Before deploying anything, the foundation needed to be solid. Tightened input validation with a minimum length check and a prompt injection guard that rejects inputs containing prompt-like keywords. Confirmed the API key is safely stored in .env and never touches the repo. Cleaned up the README with proper run instructions, project structure, and notes on Groq free tier limits for anyone who wants to run it locally. Not the most exciting day to post about but it is the kind of work that separates an app that works from one that holds up. Deployment is next ✨ Check out the repo to follow along. github.com/Durga-Anand-2006/cookable @decodingdatascience @ddsbusinesscircle #DecodingDataScience #BuildingAIApplicationChallenge #AIChallenge #BuildInPublic