A software engineering student's journey from excitement to disappointment to exposing the truth about India's "revolutionary" AI
TL;DR:
I tested Krutrim (2024) and Kruti (2025). Both were slow and error-prone; Kruti repeatedly returned identical, scripted identity replies and showed behavioral signs of being built on LLaMA 3 with heavy system prompts. That’s fine if disclosed — it’s not fine when marketed as “revolutionary.”
The Day I Believed in the Dream 🌟
February 2024. I'm scrolling through my Twitter feed when I see it - Bhavish Aggarwal, the founder of Ola, announcing something that made my heart race as an Indian CS student:
"India's own ChatGPT is here! Krutrim AI - built for Bharat, by Bharat!"
Finally! As a final-year Computer Science student, I'd been watching OpenAI, Google, and Meta dominate the AI space while India seemed to be playing catch-up. Here was our chance to show the world that Indian developers could build world-class AI too.
I was pumped. I was proud. I was about to be very disappointed.
First Encounter: The Red Flags I Ignored 🚩
Within hours of the announcement, I signed up for Krutrim's beta. My expectations were sky-high - this was supposed to understand Hindi, Tamil, Bengali, and 20+ other Indian languages better than any Western AI.
My first question: "भारत की राजधानी क्या है?" (What is the capital of India?)
Wait time: 30+ seconds
Response: Technically correct but... wait, why did it take so long for such a basic query?
I brushed it off. "Beta version," I told myself. "They'll optimize it."
The Identity Crisis That Should've Been a Red Alert 🚨
But then I asked the most basic question any AI user asks:
My question: "Who are you?"
Krutrim's response: "I am a Large Language Model created by OpenAI."
My reaction: Wait... WHAT?! 🤯
I stared at my screen in disbelief. India's revolutionary AI just told me it was made by OpenAI? I refreshed the page, asked again, same response.
My patriotic heart sank, but I rationalized it: "Must be a bug. They'll fix it."
The Questions That Broke My Heart 💔
Over the next days I ran quick tests, trivia, math, simple code. Two patterns emerged: slow responses and flaky accuracy. Examples that stuck with me:
- A wrong answer about the 1983 Cricket World Cup.
- A 15–40 second delay for simple arithmetic or a small Python function.
- Code snippets with logical errors that a CS student can spot instantly.
I felt embarrassed on behalf of the product. If this was being touted as India’s answer to ChatGPT, it wasn’t a great look.
But I'm an optimist. "They'll improve," I kept telling myself. "It's just v1."
Plot Twist: Kruti Launch and My Detective Work 🕵️
Today, August 23, 2025. I was scrolling Instagram in the campus lab when I got a notification from Ola, another announcement that made my heart skip a beat:
"Introducing Kruti - India's first agentic AI assistant!"
"Powered by advanced Krutrim 2 models!"
"Next-generation AI for India!"
Despite being burned before, that familiar excitement crept back in. "Maybe they actually fixed everything this time," I thought. "Maybe Krutrim 2 is the real deal."
I tried it immediately.
The Investigation Begins 🔍
My developer instincts kicked in harder this time. After 18+ months of studying deep learning and prompt engineering and getting fooled once, I approached this with the curiosity of a CS student and the skepticism of someone who'd been burned before.
What I saw:
Every question took 15-20 seconds to respond, Even for basic queries like “What’s 2+2?”
The Identity Question That Revealed Everything:
My question: "Who are you?"
What happened:
- Web search initiated (I could see the loading indicator)
- "Thinking..." for 15 seconds
- Final response: "I am Kruti, an AI assistant developed by the Krutrim AI team. I'm powered by Krutrim 2 and other advanced open source AI models."
Wait. It needed to do a WEB SEARCH to know who it is?! And what's with "other advanced open source AI models"?
The Detective Work: System Prompts & Model Tells 🕵️♂️
I pushed further for some follow-up questions.
Follow-ups yielded the same scripted reply every time:
I asked “What models are you using?”, “Who made you?”, and “What is Krutrim 2?” — and each time the assistant returned the identical sentence:
“I am Kruti, an AI assistant developed by the Krutrim AI team. I'm powered by Krutrim 2 and other advanced open source AI models.”
Identical, word-for-word replies like that scream of system-prompt hardcoding, not genuine model reasoning.
The Smoking Gun 🔫
After hours of prompt-injection tests and timing measurements (yes, I spent my Saturday evening on this) the pattern was clear: latency and token-generation speed matched LLaMA 3 behavior, failures were the same kinds of hallucinations and reasoning gaps, and identity queries returned a canned sentence every time. In short — heavy system prompts + light fine-tuning on an existing open-source base, wrapped in marketing.
The Technical Reality Check 💻
Let me break this down as a CS student who's actually studied how LLMs work:
What Krutrim Claims:
- Proprietary "Krutrim 2" model
- Built from scratch for Indian languages
- Revolutionary architecture optimized for Indian context
What I Actually Found:
- Base Model: Meta's LLaMA 3 (obvious from behavior patterns)
- "Innovation": Heavy system prompts and fine-tuning
- Identity Responses: Pre-scripted, identical answers to avoid revealing base model
- Performance: Still terrible - 15-20 seconds per basic question
- Identity Questions: Requires web search + thinking time
- Accuracy: Still making basic factual errors
Why This Stings as a CS Student 📚
Here's the thing - I’m pro–open-source. I build on OSS all the time. The issue here isn’t that they used open-source, it’s that they appear to be hiding it and selling it as proprietary innovation.
This matters because:
- It erodes trust in Indian AI startups.
- It wastes resources if investors believe there’s a unique, fast model under the hood.
- It distracts from genuine engineering work that could actually improve performance for Indian languages.
What Should Have Happened:
"We're building India's best AI assistant and AI agent using LLaMA 3 as our foundation, with specialized fine-tuning for Indian languages, cultural context, and use cases."
That's it! Honest, clear, and actually impressive from a product perspective.
The Reality: It's Still Just Slow LLaMA 3 (With Extra Steps) 🦙
After all this investigation, here's what I'm convinced Kruti actually is:
Krutrim's "Revolutionary AI" =
LLaMA 3
+ Heavy system prompts to hide identity
+ Pre-scripted responses for deflection
+ Some fine-tuning for Indian content
+ Web search integration (badly implemented)
+ Marketing budget
+ A prayer that CS students won't notice
Call to Action: What We Can Do 🚀
As Developers:
- Test identity questions - Ask "Who are you?" and watch for scripted responses
- Time the responses - 15+ seconds for basic questions is a red flag
- Test follow-ups - Identical word-for-word responses indicate system prompts
- Share your findings - Help the community make informed decisions
As Indian Tech Community:
- Demand performance benchmarks - Speed and accuracy matter more than marketing
- Call out scripted responses - Real AI doesn't need web searches to know its identity
- Stop falling for "Version 2.0" hype - Judge by performance, not version numbers
- Support companies that are honest about their technology stack
The Uncomfortable Truth 💭
I wanted Krutrim to succeed. I really did. Twice.
As an Indian CS student about to graduate, I dream of working for Indian companies that compete globally on technical merit, not elaborate deception schemes.
But here's what really hurts: The performance got WORSE. Krutrim was slow, but at least it tried to respond naturally. Kruti takes longer AND gives robotic, scripted responses designed to hide its origins.
This isn't innovation - it's regression with better marketing.
Discussion — I Want Your Experience 💬
- Did you test Kruti? What did you find for identity questions?
- What’s the slowest you’ve seen a basic answer take?
- CS students: how do you validate model provenance in practice?
Drop your test results and clips in the comments — let’s build a shared dataset of evidence.
PS: I’m not against Indian AI companies or building on open-source models. I’m against slow UX, scripted identity replies, and marketing that pretends something basic is revolutionary.
Follow me for more honest tech reviews, interesting tech Blogs, and the occasional rant about overhyped startups that waste our time 😤
— Puneet, final-year CS student.
Top comments (0)