Before we get started, here’s a fun exercise for you. Which of the below images is AI generated and which ones are real?
“Aaj kya khana hai? Aaj kya banana hai”? This is the one question that plagues most households every morning! There are the blessed few who are happy to eat the same thing everyday, but for the rest of us, there is a need to try new flavors, expand our palette and generally try food that is delicious and hopefully guilt free
Our friends at mazoto (or zomaito or zomaato) are trying to bring ghar ka khana to you in new and exciting ways. For those of us, whose biggest challenge is deciding what gets cooked today, we are introducing Recipe Rover.
We wanted to make instant recipes as good a feature as instant delivery. Not only that, we only wanted to show you recipes we know you will be able to make right away (because we have all the ingredients available). But creating this convenience was time-consuming to build: it took us 3-4 hours to build just one recipe page.
We were also limited by our understanding of cuisine and event-specific recipes. Regional and festival-specific recipes were an even bigger problem. Over time, there were multiple missed opportunities for us due to the high lead time and lack of context.
To solve this, we introduced the AI Recipe Rover!!
Now we have an almost unlimited repository of recipes available through our search bar along with a curated list of products that you need to cook them. Note - if we can’t help you with the ingredients, we don’t show you the recipe.
How does this help our customers?
- Simplify meal planning: Range of simple recipes with ingredients that can be quickly delivered - all in a few clicks
- Occasion-based curation: Mango season - head to the mango recipe corner; Ganesh Utsav - make modak at home
- Understand nutritional value: Each recipe also provides information about the macro-nutrients that the meal contains. Choose healthy :)
This is where our AI-enabled Recipe Rover enters the picture. It helps us generate thousands of recipes in a couple of hours - enabling us to cover a range of recipes in the context of ingredients, events, and festivals at the push of a button.
So, how did we build this entire engine?
Building Recipes: How GPT changed the game
I tried using Chat GPT to curate recipes and as a former chef, I was amazed by the quality of AI-generated recipes. We harnessed the power of OpenAI's GPT-3.5 and GPT-4 to generate a staggering 50,000 recipes at once, significantly expanding our recipe bank.
We further refined our understanding of customer preferences using our own data, allowing us to cater to customer’s unique tastes and create an ever-evolving culinary experience. Our Search team was able to take these APIs and refine prompts that could help build JSONs that could be directly parsed into our current elastic search system.
Building appealing recipes: Using Midjourney
To truly captivate our customers, we needed more than just well-written recipes. With the help of Stable Diffusion 1.5 and Midjourney techniques, we crafted visually appealing images for each recipe, perfectly complementing the AI-generated text and offering a glimpse of the final dish. These stunning visuals not only elevated the overall presentation but also helped capture the attention of our customers.. For this, we fine-tuned our own Stable Diffusion model using over a million images. After all, we had to teach an AI model what does desi chowmein look like? Or how the texture of paneer is different from tofu!
Building appealing feasible recipes: Vector search
Vector search plays a crucial role in Recipe Rover's ability to provide customers with an effective way to discover the right ingredients for their recipes.
By extracting ingredients from AI-generated recipes and matching them with available products on Blinkit, vector search ensures that customers are only shown dishes they can cook at that time. This intelligent approach to ingredient matching allows us to help customers find all their ingredients in one go!
Showcasing well-built recipes: Dynamic search & recommendation
When customers search for popular keywords, the app displays recipe image cards under the search results. These recipes are complete with ingredients to order, nutrition information, and cooking directions.
Additionally, an auto-suggest feature to help customers discover new dishes based on their preferences. We also showcase a rotating set of dishes on the homepage every day, ensuring that customers always have something new and exciting to explore.
AI + Humans >>>>> AI
A big shoutout to Shreya, Vishal, and Gaurav for their exceptional work and dedication in bringing Recipe Rover from ideation to production within a fortnight. As well as Prabljit, Saket, and Nikunj for helping us integrate it with the consumer app. Their expertise, creativity, curiosity, and collaboration were instrumental in the development and speedy release of this groundbreaking feature. We are incredibly proud of the impact Recipe Rover has had on our platform and the delight it brings to our customers. This effort also aligns with our philosophy that humans working with AI will always be much more effective than just AI or only humans.
Our team has started writing more than 50% of their code with the help of AI ;)
And to top it all - all images at the top were generated by our in-house AI model Swaad-1 (Salivating With AI-generated Amazing Dishes). No kidding.
Vaibhav Bhutani is an AI Product Lead at Blinkit. We are hiring. Feel free to connect with him on LinkedIn or drop him a line on [email protected]