The Rachna project bridges a few interesting product concepts, blurring the line between content consumption and creation.

Product Philosophy

Our Product philosophy is that

• [Loneliness / Boredom Problem]
  the world has a loneliness crisis, and the demand for connection and engagement is robust
• [Pure AI Generation does not work]
  Purely AI generated content leaves the user disengaged, in an AI-cannibalized garbage-in garbage-out world of generational content-decay
• [Purely Human Driven Engagement Does Not Work]
  purely human driven experiences leave too much work on the user, providing little value to fatigued and busy users
• [Accessibility is Key]
  the content / experiences must find the user where they already are, and that engagement strongly correlates with ease of access. This informs a choice of Telegram/Whatsapp as a key MVP platform to test our concept before getting into more serious 1st party projects.

Having read some books from the famous Goosebumps Series as a child (one of the popular picks at the school library), I couldn't shake the feeling of how engaging it was compared to books of the time. I could "choose" the outcome of the story, and make pivotal decisions through every chapter - how cool is that? There were multiple endings, replay-ability : and ultimately, high engagement. I recently ran a quick search on wiki to find it was the second highest grossing series of all time, right behind Harry Potter. Nice.

The story behind many stories

This project was born as I reconnected with Ridd in Q3 2024, with both of us bonding over our very many battle scars from our previous startup experiences. Ridd found one of my reports from Sachiv.ai and YC Founder Matching, and got curious about Sachiv, and my work on AI Agents. This sparked a fantastic partnership that has led to a handful of A/B tests (as of 2024 end, prior my US departure), a working product, and a small team. Ridd's natural patience and operational persistence compliments my vision, impatience, and technical expertise - allowing both our ideas to come to life.

Within a couple of weeks, we had a working prototype up. Within a month, we were planning our first launch. We launched the product on Telegram with some instagram advertising and amassed close to 400+ users within 2 days.

About the Product and Tech

I built the entire async stack, including an LLM powered backend on GCP (Cloud Run), accessing LLM's through OpenRouter. The choice of Telegram as an initial platform was informed by ease of deployment and simply how much we got out of the box (payments, distribution, session management, etc). There were also few overheads compared to a 1P solution on Google/Meta/Apple powered distribution channels - which require extensive setup, policy, and approvals work to kick off. The platforms (rightfully so) take great care to protect their users, albeit some of the overheads for startups (especially pre PMF) can be detrimental. While we'd support more full-fledged 1P experiences soon (I can't wait for realtime AR/VR generation), we took a call to run with Telegram to prove the concept.

It was a fantastic decision, as building the entire Telegram-AI stack took less than 2 hours on a fateful weekend in September 2024. I debated writing a custom adaptor for our persistence layer, but in true startup fashion - we made do without it for a lot longer than we had expected. Long enough, that when I onboarded our most experienced engineer to the team, he couldn't help but express (about the stack):

the comfort... is not coming...

I nearly spat my coffee in amusement.

It is a fact: the engineering traits/biases you need to go from zero to one truly differ from those valued in large-scale production environments - so much so that applying the wrong approach in a given environment is quite-often a death sentence. I speak as one who has died and been resurrected a few times over, taking down (or so I thought anyway) some large-scale production websites, and also burning my runway to the point of (almost) having to sell my house. No mistake, there were a thousand things the engineer in me too wanted to "fix" and improve pre-launch. And the PM in me wanted to launch... tomorrow. Product had to win.

The website, on rachna.live is your average Vercel / Next.JS deployment, though I take a little bit of pride in own implementation of the faded carousel of images. The images themselves though are Ridd's (painstaking) work with various free and paid AI tools. Great job, Ridd.

The innovative bit

Well, how do you get around limited context windows for LLM's? How do you make an AI assist a user "through" a story, while (1) not losing track, (2) keeping things exciting, and (3) having some level of determination in terms of outcomes (although some solutions in the market really leave things open-ended, while others offer MCQ's in a very direct way).

I described this as:

Our job is to provide the tracks for the user to drive (their go-karts) on. There's a few paths, there is complete feeling and experience of freedom, blended with guided entertainment. But, there are still some endings.

I found the answer to be in combination of LLM function calling, and multi-tiered directed acyclic graphs (DAGs), combining a few ideas from

• my time in Google Maps (you didn't think a sub 100ms response to route you from any point "A" to "B" in the world happened in a single "tier", did you?)
• my time in Google Health and Fitness, and my first exposure to DAGs in production - albeit, not a great one (that is, a story for another day...)
• and working with AI Agents for over a year

I'd love to tell you more, but you know...

Monetization

There is a big difference between a "user" and a "customer". Only well-funded startups can afford to pay for users

Our other Product challenge was how to monetize the product. Product startups are hard and take time.

We knew we can bring in users if we offer the service for free, but there is no business or opportunity unless there is willingness to pay. There are purists who would disagree with me (and rightly so) here on the principle that engagement trumps all (and boy, did we have engagement), but both Ridd and I had bled plenty paying for usage on our previous products. This is the luxury of the well-funded, and to be frank - even investors have (increasingly) limited appetite for the true product adoption lifecycle arc. Most like to step in (on a rocket ship) once the individual founders have already eaten and mitigated those risks/costs.

Investors don't pay to help you build your train. They only really want to buy a ticket aboard, once it is taking off.

Whether or not we took (or take on) investment, or they admit that it is the case - investor sentiment also heavily depends on customer willingness to pay. The ideal scenario is that the product pays for itself. I.e., we find the elusive holy grail of

and can instantly enter "scale up" mode, but even in anticipation of that - we wanted insights on what are users really willing to pay for. This, we got from the our feedback forms, and closely following the marketing and content users were responding to.

PMF : The art of bringing the product to what the market wants, AND choosing / narrowing the market to what the product truly is - until they're "close enough" for viability.

After a handful of quick iterations and A/B tests, we got our first payment 🎉. Ridd and I both treated ourselves a well-earned beer that day.