DEV Community: Puneet Chandna

🚨 I Fell for the Krutrim Hype (Twice) - Here's Why You Shouldn't

Puneet Chandna — Sat, 23 Aug 2025 16:04:13 +0000

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.

10 Must-Read Books for Software Engineers in 2025 📘💻

Puneet Chandna — Mon, 28 Jul 2025 16:28:43 +0000

As a software engineer, one thing has been constant: learning never stops.

So here’s a list of 10 books every engineer should read at least once in their career.

They’ve shaped how I approach problem-solving, systems, and even team collaboration.

✅ Books I’ve Already Read:

1. Designing Data-Intensive Applications (DDAI)

My absolute favorite. This book changed how I design systems.

I used its concepts to optimize a DB that reduced query times by 40% on a real project.

I’ve highlighted the key sections and revisit them regularly — it’s more of a playbook than a textbook.

2. System Design Interview – Vol 1

Great for both interview prep and practical, real-world system architecture.

3. Introduction to Algorithms (CLRS)

The classic. Dense, but worth it if you want to master the core fundamentals.

🕐 What’s Next?

The remaining books are still on my list, and I plan to read them one by one — as soon as I get time.

If you're building a roadmap for serious engineering growth, this list is a great place to start.

⭐ My Top Recommendation?

Without a doubt: DDAI.

It’s not just for learning — it’s for relearning.
I revisit it every few months.
It stays relevant across systems, databases, and scalability topics.
No book can beat that.

🔁 Your Turn:

What’s one book that changed how you think as an engineer?

Drop a comment, link your review, or just say hi 👋

Let’s build the ultimate software engineer reading list together!

#softwareengineering #backenddevelopment #books #learning #systemdesign #career #developer

I Accidentally Discovered a Hidden Gem for Testing Premium AI Models (Completely Free!)

Puneet Chandna — Sat, 26 Jul 2025 20:19:57 +0000

Originally posted as a LinkedIn discovery that I just had to share with the dev community.

The Discovery That Made Me Break My "No Social Media Posts" Rule

I'll be honest,I don't usually write LinkedIn posts or blogs. But sometimes you stumble across something so useful that you feel obligated to share it with fellow developers and AI enthusiasts.
That happened to me recently when I discovered LMArena (lmarena.ai), and it completely changed how I approach AI model testing and comparison.
What Exactly Is LMArena?
LMArena is essentially a public arena where AI models battle it out, anonymously. Here's how it works:

You submit a prompt
Two AI models respond (you don't know which models they are)
You vote for the better response
The results feed into a public leaderboard based on real user preferences

But here's the kicker,while you're participating in this research, you get free access to premium AI models that normally cost serious money.
The Model Lineup (And Why It's Impressive)
The platform gives you access to models that would typically require expensive API credits or premium subscriptions:

Claude Opus 4 - Anthropic's flagship model
Gemini 2.5 Pro - Google's latest and greatest
DeepSeek R1 - The reasoning powerhouse
Grok 4 - X's premium AI model
And many more...

No login required. No credit card. No sketchy popups or malware concerns.
Three Ways to Use LMArena

1. Arena Mode (The Classic)
Submit your prompt and vote between two anonymous responses. Perfect for:

Testing prompt engineering techniques
Getting multiple perspectives on coding problems
Comparative analysis without bias

2. Direct Chat Mode
Choose a specific model and have a direct conversation. Great for:

Deep-diving into technical problems
Iterating on code solutions
Model-specific testing

3. Side-by-Side Mode
battle between 2 models of your choice, great for:

Understanding model strengths and weaknesses
Choosing the model for your use case
Research and analysis

Why This Matters for Developers

Cost Savings
Instead of paying for multiple API subscriptions to test different models, you can evaluate them all in one place for free.

Unbiased Comparison
The anonymous voting system removes brand bias. You're judging purely on output quality.

Real-World Performance Data
The leaderboard reflects actual user preferences, not just benchmark scores.

Prompt Engineering Laboratory
Perfect environment for testing how different models respond to various prompting techniques.

A Word of Caution (And Why I Trust This One)
I've seen countless "free GPT" clones online, and most are either:

Spam-filled nightmares
Potential security risks
Barely functional wrappers

LMArena is different because:

It's research-backed and transparent about its purpose
No data collection beyond the voting mechanism
Open about its methodology and model selection
Clean, professional interface without dark patterns

Real-World Use Cases
Here are some ways I've been using LMArena in my development workflow:

Code Review and Debugging
Prompt: "Review this Python function and suggest improvements for performance and readability:"
Getting multiple model perspectives helps identify issues you might miss.

Architecture Decisions
Prompt: "Compare microservices vs monolithic architecture for a team of 5 developers building a SaaS platform"
Different models often emphasize different trade-offs.

Documentation Writing
Prompt: "Explain this API endpoint in simple terms for junior developers"
Comparing explanations helps you find the clearest communication style.

The Bigger Picture
LMArena represents something fascinating in the AI space,a democratized testing ground where models are evaluated based on real user needs rather than academic benchmarks.
As developers, we often need to choose between different AI tools for our projects. Having a neutral space to test and compare these models without financial commitment is invaluable.
Getting Started

Visit lmarena.ai
Choose your mode (Arena, Chat, or P2L)
Start testing with your own prompts
Vote on responses to contribute to the community

No signup, no credit card, no commitment.

Final Thoughts
I'm sharing this not because I'm sponsored (I'm definitely not), but because good tools deserve to be known by the community that can benefit from them.
In a world where AI access is increasingly paywalled, LMArena feels like a breath of fresh air—a place where you can experiment, learn, and contribute to AI research simultaneously.

Just remember:

if this tool proves as useful to you as it has to me, consider sharing it responsibly. Good free resources tend to get overwhelmed quickly, and we want this one to stick around.

Have you tried LMArena? What models performed best for your use cases? Drop your experiences in the comments below!