Manas J. SaloiModify

2025-03-02 08:00:00

Behind the Plate - Swiggy Originals Series

Swiggy should do a series called “Behind the Plate” where they showcase the actual people who make your food. Not those polished chef interviews you used to see on TV, but something raw and authentic. There’s this huge emotional connection with food that most apps completely miss out on.

The idea is simple. Each episode features one chef from a restaurant you order from. Make the chef the hero of the story.

Think of this less like Chef’s Table or Nikhil Kamath inviting restaurant owners to discuss French patisserie techniques. The goal, the audience, and the format are different here. Chef’s Table tells the story of restaurant owners perfecting their craft—not the young cook from UP who moved to Bengaluru with big dreams and now supports his family back home by making chole kulche. This will focus on the latter.

No script, just real conversations about their journey, struggles, and what drives them to cook.  Show them actually making their signature dishes. Let viewers see the care that goes into the food they order with a few taps on their phone.

I remember eating at this small restaurant in Ubud. It was run by this one old lady with her daughter. The menu had this story about how they started. How they care about the ingredients and how they source them. Made me appreciate the food even more knowing the story behind it.

There are three things this kind of content could do for Swiggy:

• First, it builds trust. When you see the actual people making your food, their dedication to hygiene and quality, you’re more likely to order (hypothesis). Food delivery has this trust barrier that’s hard to overcome. With adulteration cases going up, people have started thinking more and more about what they put inside their body. This will also show the kitchen, and implicitly signal how hygienic the place is.
• Second, it creates an emotional connection. We all love stories. When you know the chef spent 10 years perfecting that biryani recipe that they learned it from their grandmother, you’re more invested in trying it.
• Third, it differentiates Swiggy from every other food delivery app. Zomato, Swiggy, they’re all starting to look the same. Features get copied within weeks. In a space where technology is quickly commoditised, the stories you tell could be a differentiator. This kind of content library? That’s harder to replicate. Keep it simple, authentic, conversation based. The rough edges make it feel more real. These will be Instagram reels and Youtube Shorts. Restaurants will be happy to support this as Swiggy will be putting in the effort to market them.  Vibe of the shorts will be a mix of fave.finds.blr’s Koshy restaurant reel + Ranveer Brar’s Kolkata street food reels.

Once you have enough content, you can even showcase this in the restaurant page. Will this directly increase orders? Maybe not immediately. But it changes how people think about Swiggy. It’s no longer just an app, but a connection to real people making real food. Every order becomes a story, not just a transaction. I don’t know if Swiggy is thinking about content as a moat, but they should be. (Another similiar idea I posted was live streaming the inside of a kitchen to build trust. This was during peak covid. That never took off. Maybe this one will.)

Added context:

People on Twitter were commenting saying this is Chef’s Table in replies. Another series by Zomato. Nikhil Kamath has done episodes with other restaurant owners too.

But this is not that.

This is not about the owner of a restaurant chain who is telling you how he got a loan to open his first restaurant. This is about the small-town guy from a tier 4 city who moved to Bengaluru with dreams, has a family to support back home, and is now making chole kulche for you at a restaurant with a colorful name like “Ustad ji ras bana rahe hai aur kofta khila rahe hai.” He tells how he adjusted the chole kulche recipe to make it feel like home.

Even better would be a north Indian cook making dosa at some darshini.

It is reels. 1-minute videos. Bingeable content.

Chef’s Table is about how a Michelin star restaurant owner grew his own potatoes. You can’t show this level of detail in a YouTube short or reel. You need an hour. It is niche and not for the mainstream audience.

This is for the masses. People tagging others and saying, “Oh I know this small town where this cook came from. It is close to where I grew up.” It is about nostalgia. It is about connecting with others.

Showing the restaurant kitchen is of course an important but subtle part to signal hygiene. But the storytelling has to be stellar.

The vibe is important. See Ranveer Brar’s Deckers Lane reel. There is this other Pakistani Instagrammer I had seen. She would go around Karachi eating at small street food places. This is mostly like that, but also focuses on the cook.

2025-03-01 08:00:00

AI can:

• Help ICs do their current work:
  • Better.
    • For a PM, can their spec become better because now they can use tools like Deep Research (OpenAI) for competitor analysis, Deep Search (Twitter) for faster, efficient search? Can AI give insights about customer behavior and point out edge cases that a normal PM would miss?
    • Can a copywriter now come up with dozens of copy variants that are probably far better than anything they could create on their own?
  • Faster.
    • Lead time: Can you cut the lead time to arrive at a spec and prototype? This is easier because you can leverage existing templates, turn notes into first-cut specs faster, and brainstorm with LLMs to think through ideas and stress test them. As a PM, you can work on quick prototypes before you collaborate with a designer on the solution. These will mostly have to be discarded. Tech leads can create their tech specs faster. AI won’t do the complete work for them, but previously the planning stage would take a lot of time. Now it is far faster to brainstorm, go broad, and then narrow in on a solution (diamond).
    • Cycle time: Here it is trickier. Landing pages for marketing would be far faster. Marketers can probably do these on their own now. Stand-alone pages. You can create copy faster. If midway through the project execs change their mind, instead of the project being on hold until designers change their designs or add more scope, you can quickly use the designs to create a web prototype, make the changes, and show it, instead of making changes on the actual app. [I wish this did not happen, but ask any company, and they will tell you scope changes and design changes, especially in startups, happen quite a bit.] For a senior engineer, tab tab tab works because even in a big codebase, you are not using the agent to make changes across multiple files. Scaffolding is taken care of. Just don’t expect some 90% improvement though. It is not going to happen unless you want to go full Yolo mode and spend months refactoring later.
  • More.
    • If you are doing better and faster work, a person is already delivering more. So they can manage more projects at a time and also take less time per project.
    • Since you can generate potentially infinite variants of copy and designs, you can create far more marketing campaigns than before. You can try many more landing pages. You can test different funnels for different segments. Marissa Mayer used to test 40 shades of blue for a button. AI gives you the same power, without needing the resources of a Google.
    • Empower an IC to take on completely new work not possible before:
      • This is the most exciting part of this AI age.
      • A marketer is now a copywriter.
      • A PM, instead of waiting for a data analyst for complex queries, can do their own analysis.
      • A marketer can create designs for their campaigns using various design tools.
      • A PM can make prototypes while collaborating with their business counterparts or engineers, leaving the designer to do deep work on their current projects.
      • A backend developer working on an internal dashboard does not need to wait for a designer. They can use v0 for it. They don’t even need to wait for a front-end dev. They can use v0 again. Do note: These tools are good for standalone projects with smaller code bases. You can’t expect these gen AI tools to write scalable code, and as the code base increases, it will hallucinate more and might even want to delete your own codebase for the lols.
    • Help teams coordinate better:
      • There are tons of meeting notes and action item products. No meeting should end without a Minutes of Meeting (MOM) now.
      • People hate writing. An email takes effort. Even a Slack update requires work. Superhuman has autocomplete. Gmail has summaries. You don’t need to wait 3 days to reply to an email going forward. Replies might be templated and get shorter, but people will not become bottlenecks (hopefully).
    • Enable Push vs Pull:
      • This is my favorite topic and something I have written a lot about.
      • Push »»»> Pull.
      • All PMs pretend to like data. But no one opens Tableau every day to check data. An email update means you are not pulling data anymore, and it is delivered to your inbox. Earlier you would wait for a data analyst to create a new email, but maybe now you can do it yourself. Or the data analyst can be more productive and create a new insights email and not need a week to get it done.
      • What are the insights people need to do their job better? How can we push them in a timely manner with clear action items? Facebook campaign performance is declining? Is there an email with updates about that, with insights on why that might be happening, along with 5 more suggested campaign ideas?
    • Push 10X more content on your distribution channels:
      • Use your distribution to your advantage. Create more AI-generated videos. Create memes. Post replies on other channels. Maybe create a completely new channel with only AI videos. More thoughts:
      • 90% and not 100%:
        • The more specific the output, the harder it is for AI to get it done. It is a next token predictor after all. This is why I can come up with a dozen ideas for a wedding card (yes, I tried this for a friend), and the characters can seem like them, use style and character reference, but for some reason or another, the hands will be messed up on Midjourney. Yes, I can use Replicate and try out more models, but each incremental rerun costs money. This is what companies will learn too. So the best use cases are where you don’t need 100% accuracy, but are okay with 90%.
          • You might want creative ad ideas.
          • You might want creative landing pages.
          • You might want nice graphics.
        • But be happy with 90%. If you try to make it 100%, you might end up spending as much time as before on Photoshop and you won’t get the productivity gains you were expecting.
      • Workflows:
        • I would do an audit of workflow of each person working for my company and understand how they can leverage these tools to do more work at a faster pace, and can take on adjacent work.
      • Existing tools vs in-house tools:
        • I tried this experiment a while back. I was trying to see if I can create a MiniManas, use AI to write like me. I have 500 blog posts on my site.
          • I started with just asking all the chat tools to write like me without example.
          • I then tried to use better prompts.
          • Asked ChatGPT to update its memory on how to write like me. Used Claude’s styles.
          • Then I gave an example in the prompt. Gave some notes.
          • Then I tried to give examples of drafts vs final output, and played around with the number of examples and final output.
          • Then with the help of a friend, I fine-tuned a model and tried the same exercise.
          • Next would have been to use some RAG-based system to create embeddings of all my written content, chats or blogs, store them in some vector store, and then use different LLMs to write like me. I tried the first part, but the embeddings workflow was failing. And I was too poor to use OpenAI’s embeddings API.
          • I gave up at this point.
        • My point with the above is that you can try all sorts of solutions. But instead of trying to fine-tune some LLM and creating your own workflows, see if you can use an existing tool. You don’t need to fine-tune some LLM to write code. You are not JaneStreet who needs Ocaml and needs their in-house tab tab tab. You can just use Cursor. Yes, you might have a particular design system and brand guidelines, but that does not mean you need custom loras to generate images for a brand campaign. You can just try prompt engineering and get 90% there. Post an existing brand image on ChatGPT or Grok. Ask them to describe the image. Use the description and reference image to create more brand images on Midjourney. Take it to Magnific AI to upscale. You can even use a mix of Midjourney + Figma. Instead of creating custom in-house agentic workflow, maybe try some off-the-shelf available agent workflow builders.
      • I particularly love what a founder told me: His goal is to make sure every employee in the middle of the bell curve gets superpowers and can do their job like an A player. I think that is a nice way to think about this transition.

2025-02-28 08:00:00

Uber’s business model change for Auto does not matter. It is too little, too late.

Every strategic move is offensive or defensive. Rapido launching 0% commission and charging flat platform fee from drivers was an offensive move.

If an incumbent has fat margin structure, you can attack its margin. Rapido makes money from 2W. So this expanded use cases for Rapido. Gave customers more choices. Supply switched to Rapido.

Even this was a response to Namma who did this first.

Assuming since then a lot of supply has moved to Rapido and Namma, and demand has followed supply to these 2 platforms, then this change from Uber is meaningless.

It won’t attract demand back to Uber. I don’t remember when was the last time I used Uber. Uber does not even have differentiated supply. Cars have become worse. No quality control issues.

Support is bad. Support is bad on all of these platforms, and it means for an user there is no clear difference between any of these platforms.

For a premium user there is no need to choose Uber over Namma or Rapido.

This commission/ business model change should have been made much earlier. But they could not. if they had changed the commission structure only for autos, their car drivers would have revolted.

And even if their India business revenue was only 1/2% of global, for a public company it was still meaningful. So they needed their revenue from cabs. 25% take rate is meaningful.

My guess is that Rapido and Namma took market share away from them. Not just in auto, but also Cabs. Rapido is 0% commission even on cabs.

This is a last ditch defensive attempt from Uber to protect GMV (assuming auto rides on Uber’s platform is significant % of their total rides), and give up on revenue.

And Dara made it clear that he is focused on only countries where they can be clear number 1 or 2.

So India will never be a focus for them. Eventually they will move to 0% even for cabs to compete, and run this business with a far leaner team. Just a technology player providing platform service and not a highly managed marketplace.

2025-02-27 08:00:00

We have spent the last couple of years optimizing funnels. We had to. When your competition has 4X your cash balance, you have to learn to optimize and get more out of everything: your user, your funnels, and from each transaction that happens on your marketplace. How do you think about funnel optimization? You look at each segment and their journey on your product. New users will always have lower conversion than a mature product. So they need to be treated differently than a mature user. The interface has to adapt to each segment.

A lot of people talk about dynamic interfaces on Twitter. You can take the idea to the extreme: one-time disposable interfaces, on-the-fly generated UI created by AI. But familiarity is important. No user should have to think where the call-to-action button is every time they come to your app. Too much cognitive load for the user. As with most things, the answer is somewhere in the middle. Take all user segments. For all, destination search would be key on the homepage of a ride-hailing app. But the size of the search can change. For a new user, you can use some animation to bring focus on the destination search button. You can hide everything that distracts the new user. Yes, if the user has come from a Facebook ad for rental and then lands on your home, you might want to highlight rental service on the Home Screen, but otherwise you need to make it less prominent vs search. Assume they have come to use your base service.

What matters to the new user? Call-to-action is of course search. They have not even taken a new ride. So they don’t need to have a saved addresses section. There is no one-tap booking because you can’t predict their destination, pickup, service, price, and add-ons. You need to show that they are getting a good deal. So new user vouchers need to be auto-applied. You need to give gratification to the user. Maybe some social proof that they are making the right choice (without disclosing sensitive data, not everyone can be like Zomato telling how many orders they are doing daily).

Once they have taken their first ride and are familiar with your flow, you can start progressively adding more elements on home. An adjacent point: The same component can adapt to do different things too. A card on home can let you do one-tap/quick booking. It can be continue where you left once you drop off and come back to the homepage. It can be return trip booking for daily office commuters. The component adapts based on the use case. Over time your mature users would want you to predict their behavior. Want you to cut the number of steps. Then you focus on power user needs. As I said: One-tap booking, one-tap return booking, etc. can cut the time to book the journey to 1/3 of the time normally needed to book a ride. Now if you have a user segment service that can return the segment to the client, then you can make the whole journey dynamic. Based on action taken on home, your price estimate and service selection screen can change. You can change what you show on finding driver screen. In order summary screen, you can show different components. If it was the user’s first ride, you can ask them to save their destination once the ride is complete. Show gratification around money saved through the voucher. Even cross-sell other services if you are a super app.

The interface does not need to change based on just the user need. It can be business-driven too. Our service selection screen (that I will call estimate screen going forward) is based on a slot system, basically think of everything on the estimate screen as fitting into slots. You can dynamically change the number of slots. And also information on each slot. And each slot can handle different data types.

The image for the service type can be a gif. Instead of showing details about the service, you can have info around how many people that service type can handle. Everything is fluid. Service ordering can be based on surge, time of the day, previously service taken by the user, popular service in that area, new service you want to promote, trip distance, price, ETA, or any business metric you want to optimize. And you can change anything you want from a config service. No back-end service deployment. No client release.

Now, we are the only ride-hailing company in the world that provides multi-modal trips: by multi-modal I mean you can select services for different legs of the journey and book one single order. Services can be in-house or external. You can take base ride service for first leg of the journey to metro, middle mile metro ride (we have partnerships with public transportation services), and then final leg on ride economy. So our service selection screen needs to handle multi-modal too. We should also be able to upsell all 3 legs. Imagine the complexity. I can go on and on, but hope this gives you an idea around how to think about different funnels and how your interfaces can adapt to their needs (as well as your business).

2025-02-26 08:00:00

I saw a tweet on the taste of Indian founders.

As I wrote in a post earlier, it is one of the most overused words today and everyone thinks they have it.

Fold Money probably has the best onboarding experience anywhere in the world. You can compare globally. We see Family (now part of Coinbase) and think “wow, impressive taste.”

Fold Money does not get the recognition because constantly online valley taste makers will never get their hands on it. Swiggy and Zomato can compete globally with any on-demand app on taste.

Cred has “taste.” Yes, you might disagree on some of their design decisions. You can discuss UI vs UX. But I know people there. They spend weeks perfecting small details people won’t even notice.

Zerodha products are well designed.

I could list many more examples.

We don’t have Deep Research because of other reasons and not “taste.”

2025-02-25 08:00:00

I really like the vibe of Nous Research. I collect different design styles out of interest.

I loved this site that has a collection of neobrutalism-styled components based on shadcn.

But I am lazy. I am never going to spend time on something like this, building it from scratch.

Then I saw Pietro share how he created a “glass-like” design system in one shot, with all the components. I got super excited and tried to do the same thing with Grok-3, Claude 3.5 Sonnet, and O3-mini.

It seemed boring to copy something already done. So I started asking myself what style would be cool. Can I get inspired by Mughal architecture? Can I try some Persian style? Maybe some Japanese style?

After a long discussion with all these LLMs, I came up with a few design systems, but none were as cool as the neobrutalism site. I will probably keep trying these experiments until I come up with something interesting.