Kosmo

Kosmo

Stop babysitting AI. Start creating.

Inspiration

When ChatGPT launched in November 2022, I made a decision that shaped everything after it. I would build 100% with AI. Not as a side experiment. As my primary method.

Every project I have shipped since then was built that way. Hardware prototypes, delivery MVPs, browser concepts, an AI operating system, gym platforms, cafe systems, a real estate app, a gamified learning product, an autonomous betting agent. All solo. All with AI.

But the more I built, the more I noticed a pattern. The AI was never the bottleneck. I was. Specifically, the way I talked to it.

Every project started the same way. I would describe what I wanted. The AI would build something close but wrong. I would spend hours correcting assumptions I did not know I was making. The model did not lack capability. It lacked context. It lacked boundaries. It lacked the things a domain expert would have said out loud before writing a single line of code.

That is the real problem with AI right now. Not that it cannot build. It can. The problem is that the person asking does not know what they do not know. And the AI is not going to tell them.

AI labs keep shipping more powerful models. They also keep publishing massive documentation guides on exactly how to use them. Anthropic's Applied AI team gave away forty prompting techniques for free. OpenAI published a plain four-part framework straight from the team that ships the model. These are the people who understand the tools best, handing you the manual for free.

And almost nobody reads it.

Generative AI reached roughly 53% of the world's population within about three years of its mass-market launch. Faster than the personal computer. Faster than the internet. But only about 21% of people use AI daily. Adoption is not the same as fluency. The world has adopted AI. It has not learned how to talk to it.

The gap between what a tool can actually do and what the average person knows to ask it for is not closing on its own. It is structural. It will keep existing for as long as documentation moves faster than people can read it.

That gap needs a compiler. So I built one.

What it does

You describe what you want in plain English. Kosmo asks the questions a domain expert would ask. It checks current official documentation. It injects guardrails, stopping conditions, and hidden reasoning loops. Then it outputs a structured, tool-aware prompt specifically adapted for the AI you are using, whether that is Claude, Gemini, Cursor, or ChatGPT.

Kosmo is not another AI wrapper. It is not a chatbot. It does not replace Claude or ChatGPT or whichever tool you are actually going to run your task on. It sits in front of all of them. You tell it what you want the way you would explain it to a colleague, not the way a prompt-engineering course tells you to phrase it. Kosmo does the part almost nobody has time for: it knows what the target tool's own current documentation says it expects, and it compiles your intent into exactly that shape.

You do not need to know that Claude wants XML tags and ChatGPT wants markdown headers. You do not need to know the goal-context-constraints-done-when framework OpenAI just published for Codex. You do not need to have watched Anthropic's workshop, or tracked which framework's API changed this week. Kosmo already read it.

You just say what you mean. And the tool you are sending it to gets exactly the input its own makers say gets the best result out of it.

How we built it

I built Kosmo 100% using AI. Ever since I started building in 2022, I have built with AI. For Kosmo specifically, this is the stack:

AI Development Tools:

Google Antigravity (primary development environment, used for the majority of the build)
Claude Code (secondary development environment)

LLM Infrastructure:

Google Cloud Vertex AI for the LLMs powering Kosmo's internal agents (Gemini 3.1 and Gemini 2.5 Flash)
Kosmo is an AI-first product. The entire app runs on a 5-agent mesh/harness collaborating and working together for every single user request

Knowledge Vault:

Supabase for the Knowledge Vault, a vector database where we keep the embedded, up-to-date documentation of top AI tools, important technologies, and frameworks
We customized an open-source project called FireCrawl and created a Knowledge Scraper which scrapes, formats, and embeds based on a specific source list that I personally curated
This knowledge is maintained and kept up to date with the updating documentation. The agents use it to stay current

Application Infrastructure:

Supabase (separate instance) for the application database
Clerk for authentication
Razorpay for domestic payment processing (India)
Lemon Squeezy for international payment processing
Resend for billing-related emails
Loops for welcome, onboarding, and update emails
Vercel for deployment

With all of this and about 70 days of day-and-night work put together, I finally have the first version live at askkosmo.com.

Challenges we ran into

Coming from a humble background, resources were my biggest concern. All the technologies and the tech stack I mentioned above are expensive.

But I made it work with generous free tiers. Most of the services I got through the GitHub Student Developer Pack. I started multiple free trials on multiple accounts. I could only gather enough money to buy the domain name and one month of Claude Pro.

The 5-hour daily usage limit on both Claude and Antigravity was never enough to build a system of this scale. I was constantly rationing AI time, planning exactly what to build in each session so I did not waste a single minute of the window.

When the time came to integrate payments, I found out Stripe was invite-only in India. So I had to get domestic approval from Razorpay to accept payments. And no payment aggregator allows you to apply or accepts your application unless you have a working product, which I did not have 10 days before approval. But by the time they reviewed the application, I did.

On August 6th at 5pm, the same day we launched, I got an email that changed how I thought about the rest of this build. We got accepted into the Google for Startups Cloud Program. It came with $2,000 in Google Cloud credits, backing the exact infrastructure Kosmo runs on.

Then on August 11th, 2026 at 17:45, our international payments got approved too. Lemon Squeezy accepted us. So Kosmo could finally charge users outside India as well, not just domestically through Razorpay.

Because of all the generous free tiers, free trials, and extremely helpful programs like the GitHub Student Developer Pack and the Google for Startups Cloud Program, combining all of that together, I was able to build Kosmo and beat all odds in 70 days.

Accomplishments that we're proud of

With extremely low resources, I built what I had in mind into an actual product.

The first version is working and solving the problem I meant it to solve.

I launched on August 6th. Eleven days later, on August 17th, we crossed 167 total sign-ups. Almost all of them, we connected with and onboarded manually, one by one, and 14 of them actually paid.

I built this alone. 100% using AI. With the help of multiple AI agents across everything.

What we learned

I learned how to keep extremely high signal and suppress noise while building. When you have limited AI time and limited resources, you cannot afford to explore dead ends. Every decision has to be sharp.

I learned how to produce more with less. Not as a motivational phrase. As a survival skill. Every tool I used was either free, trial-based, or squeezed from a student program. Building Kosmo on this budget was the hardest thing I have ever done. And it somehow worked out.

I learned how to actually talk to users and collect feedback. Not in a theoretical "lean startup" way. In a "the product is live and people are using it and I need to listen" way.

What's next for Kosmo

This is just the first version. It exists for people to understand what value Kosmo provides and what problems it solves, combined with what I could build in 70 days.

I took 70 days because I was building something foundational. Something that could be scaled.

The next version of Kosmo eliminates the remaining friction as well. The user will give their raw intent, answer a few questions, and Kosmo will use the AI tool (like Claude) with all the expertise and directly give the user a final result. No friction. No 100 unnecessary iterations just to get your intent made correctly. You can directly judge if the final output is what you envisioned, or if it is even better, or if some things need to be changed.

No months of research. No need to create multiple prompts manually. No constant supervision and no domain expertise required. You say it in plain English and it builds that correctly using your target tool. No middle man needed.

The human actually becomes the intent decider and output qualifier rather than an iterating, supervising middle man. Changing the definition of "human in the loop" and a human's role in using AI to its full potential.

Built by Harshal Dorlikar, Founder of Kosmo.

Try it free at askkosmo.com.