This is a transcript of my talk for the AI hackathon at Kalvium. Install Promptrepo and use this Google Sheets to follow the talk. You can also watch this video to understand the demo better

What to build?

Hello everyone, welcome to the AI hackathon! Before we dive into AI, let’s discuss about what you should build for your hackathon. There are two kinds of applications you can work on: consumer applications and business applications. Since most of you haven’t worked in business, it makes sense to focus on consumer applications. The key is to solve a problem you face every day. This approach will make your projects relatable and impactful.


Let me share an example from my own experience. Several years ago, a CIO of a British company took me out to an Indian restaurant. They served these delicious pakoras, and he loved them so much that he ended up eating only pakoras for dinner. I sat there wondering whether I should say something about how unhealthy that was. Years later, I fell into the same trap. I only liked french fries and pastries when I moved to France. If we’re picky eaters and don’t know how food is made, we tend to pick the unhealthiest options. Ignorance is bliss, right? Or so I thought until my health suffered.


Finally, I found NutriScore—a scale that evaluates the nutrients in food. Instead of showing a number, it uses a simple, intuitive scale: "A" for very healthy, "B" for healthy, all the way down to "E" for unhealthy food—like a 5-star rating, but for health. However, NutriScore was only available in France and neighboring countries in Western Europe. I wanted it everywhere—whether I was in the US or India. And that’s how Neartail - food search was born—a search engine to find healthy food near you. When you search for "Protein salad," you see where it’s available near you and can check its NutriScore by clicking on it. So, think about your own daily struggles. Whether it’s finding healthy food options, staying organized, or improving productivity, solve a real problem you care about. That’s the best way to approach a hackathon project (Also read How to get startup ideas).


How to create?

Now, let’s get into how we can use AI to solve problems, like we did with Neartail - food search:


Use Case 1: Search
The first AI use case we’ll cover is Search. There are two ways you can build a search engine:
  1. Store all the web pages in a relational database. Then, when I enter "Protein salad" in the search box, you can use a "LIKE" query to search for that word in the database:
    SELECT page_title, page_content
    FROM web_page
    WHERE page_content LIKE '%Protein salad%';
    That’s how old search engines like Yahoo used to work. In our case, "Chef Salad" wouldn’t show up in the search if we had used a "LIKE" query. Even if it is packed with proteins, it won’t show up because the phrase 'Protein salad' doesn’t appear on the 'Chef Salad' page. This is how SEO was manipulated in the early days of Yahoo—website owners would add numerous keywords to their web pages to appear in Yahoo's search results.
  2. To build a good search engine, you need to understand the meaning of the web page and the search term, then check if the meanings are similar. We use AI (embeddings) to find the meaning for pages like "Chef Salad" and the search term "Protein salad." Then, we find the closest web pages and list them. This is how modern search engines like Google and Neartail - food search work. This is why we see "Chef Salad" when we search for "Protein salad."
This method ensures that search results are not only relevant but also personalized based on context, like losing weight or building muscle mass. (Also read How AI brought ≈ to computers).

Use Case 2: Classification

Ok, we’ve listed the food based on search. Now, we have to calculate NutriScore based on nutrients, right? Not so fast—NutriScore isn’t calculated the same way for all food. The fat in avocado is very different from the fat in beef. One is healthy, and the other isn’t. So, the NutriScore calculation changes based on the category of food—vegetable or meat. Before we can calculate a food’s NutriScore, we first need to classify it. Foods fall into categories like General foods, Red meat, Beverages, or Cheese. Our classification model uses web page titles and content to identify the correct category. Let me show you an example of how we classify foods and calculate their health scores.


[Demo: Show Coke being classified under ‘beverages’ and how its NutriScore differs from ‘Chef Salad.’]

This classification ensures that the correct formula is applied—a crucial step before we even calculate the NutriScore. Once we have the category, we need to extract nutrient values to compute the score. That brings us to Extraction.


Use Case 3: Extraction

You’ve probably seen nutritional labels on food packages, right? Calories, sugar, fiber, fat—they’re all present on web pages too, but mostly as tables or paragraphs. To calculate NutriScore, we need these values in a structured format. That’s where AI steps in.


At Neartail - food search, we use AI to extract structured data from web pages. Let me show you how we input the title and content of a web page and extract information such as:
  • Ingredients (like sugar, salt, or flour);
  • Nutritional values per serving (like calories, fat, or protein);
  • Serving size (grams or milliliters).
[Demo: Show an example where a product’s serving size, ingredients, or nutrition data is extracted and generated if incomplete.]


Once we have this information, we can calculate NutriScore for each category using programming languages like Python or JavaScript.

Use Case 4: Generation

There is one problem though. Many restaurants don’t provide nutrition details for their dishes. These restaurants are often small, run by two or three people who don’t have time to maintain detailed websites. What do we do in such cases? This is where Generation comes in.


AI has access to vast amounts of data and knows how various dishes around the world are prepared and their typical nutrition values. We can use AI to generate missing nutrition data for dishes like ‘Chef Salad.’ While the output may not match the exact preparation method of a particular restaurant, it provides a reasonable estimate that can be displayed with a disclaimer: Nutritional values are estimated by AI.


[Demo: Show how the system generates guessed values for Coke and outputs complete data for an incomplete product.]


This capability ensures that even with incomplete inputs, users get meaningful, actionable insights. By bridging data gaps, we empower both consumers and businesses.


Next Steps

To wrap things up, AI is transforming the way we solve problems, from finding the healthiest foods to improving our daily lives. The four use cases we explored today—Search, Classification, Extraction, and Generation—are just the beginning. On top of these technologies, you can build agents that know your budget, find the healthiest food within it, and automatically place orders for you. Big companies and startups like ours are building such agents as we speak. So, as you work on your hackathon projects, I encourage you to think about how you can use these AI techniques to tackle real problems—whether it’s improving health, education, or anything else that matters to you. 


Once you have built a full stack solution that uses one or more of these usecases, there will be initial round of selection on Feb 1. In that round, 20 teams will be selected based on these 4 criteria (25% for each):

  1. Idea - You can choose your own problem statement. How good it reflects your own pain point is the main criteria.
  2. Solution - Which of the 4 usecases are you using and how well you have implemented them.
  3. UX - How easily can end users use your solution and how well designed your UI is.
  4. Presentation - How clearly are you presenting your project and how well you are answering our questions.

The final round will on Feb 8 & 9. If you have further questions, post them below. Best of luck to all of you!