Rishi Kulkarni

CSE (AI) Student | Building AI + Cloud Projects in Public
Java DSA • AWS • Full-Stack
Shipping projects, learning in public & documenting the journey!

[Rishi Kulkarni @TheGeeksGuy]

Feels heaven! 🤩
@HdmcHubliDwd @WHO @Hubballi_Infra @JoshiPralhad @MTenginkai
#CleanAir #HealthyLifestyle

[Rishi Kulkarni @TheGeeksGuy]

@softwareduniyas Link

[Rishi Kulkarni @TheGeeksGuy]

@abhijitwt Nobody

[Rishi Kulkarni @TheGeeksGuy]

Repository:

https://t.co/6gO6cmJoZU

Feedback is always appreciated.

Next I'm planning to explore CUDA matrix multiplication, convolution kernels, and GEMM optimizations.

[Rishi Kulkarni @TheGeeksGuy]

Final benchmark:

Dataset: 10M elements
CPU: 12.3 ms
CUDA V6: 0.142 ms

≈87× speedup over the CPU implementation on my hardware.

[Rishi Kulkarni @TheGeeksGuy]

The biggest lesson wasn't CUDA syntax.

It was learning how memory access patterns, synchronization, and warp execution determine performance.

GPU programming is fundamentally about understanding the hardware.

[Rishi Kulkarni @TheGeeksGuy]

Along the way I explored:

• Warp divergence
• Shared memory
• Bank conflicts
• Loop unrolling
• Warp shuffle
• Cooperative Groups
• Grid-stride loops
This project became much more than "parallel reduction."

[Rishi Kulkarni @TheGeeksGuy]

Instead of jumping to an optimized kernel, I implemented every major optimization step separately.

V1 → V7 shows exactly how each optimization improves the previous implementation.

[Rishi Kulkarni @TheGeeksGuy]

The experiments were run on an RTX 3050 Laptop GPU (GA107).

Understanding the hardware is just as important as writing the kernel.

• 16 SMs
• 2048 CUDA cores
• Shared Memory
• 192 GB/s peak memory bandwidth

[Rishi Kulkarni @TheGeeksGuy]

It took 7 CUDA kernel versions to understand why "adding numbers" is one of the most important optimization problems in GPU programming.

I implemented CUDA Parallel Reduction from scratch, progressing through 7 stages inspired by NVIDIA's optimization techniques.

🧵👇

[Rishi Kulkarni @TheGeeksGuy]

@markproduct Cursor has banned me from using it

[Rishi Kulkarni @TheGeeksGuy]

@striver_79 Yeah, on the other day I ordered food from @zomato the food was pretty bad, raised the issue these guy have this weird ai that kept saying can't refund, even when the food was genuinely undercooked and bad.

[Rishi Kulkarni @TheGeeksGuy]

@kirat_tw True that

[Rishi Kulkarni @TheGeeksGuy]

Worked on neural collapse for a few months now..
Project, research publication and insights coming soon!
Stay connected and discuss freely in the comments.

[Rishi Kulkarni @TheGeeksGuy]

@vicky_grok Y would anyone take it down? Its freely available on yt. On Stanford official channel

[Rishi Kulkarni @TheGeeksGuy]

What if industrial equipment could tell you it was about to fail?

I built AssetSense, an IoT predictive maintenance system using ESP32, MQTT & AWS to monitor vibration, temperature & current in real time.

My first AWS Builder article!

https://t.co/A7Yab7ILon

#AWS #IoT #DevOps

[Rishi Kulkarni @TheGeeksGuy]

The deepest form of slavery is the hunger for being understood.

[Rishi Kulkarni @TheGeeksGuy]

Interactive visualizations for:

• Transformer Attention
• Neural Networks
• RAG
• MCP
• Binary Search
• Recursion
Built with Python, Streamlit, FAISS, NetworkX, Ollama, Whisper, gTTS & Plotly.
🔗 https://t.co/Q35KHlPLfg
#Python #AI #GenAI #BuildInPublic

[Rishi Kulkarni @TheGeeksGuy]

Built Synapse – Adaptive AI Tutor at Gen AI Hackathon 2026!

An AI-powered learning platform that identifies knowledge gaps, adapts explanations to learner proficiency, and visualizes complex concepts interactively.

🧵👇

#AI #GenAI #EdTech #buildinginpublic

[Rishi Kulkarni @TheGeeksGuy]

Instead of teaching everyone the same way, Synapse:

✓ Detects knowledge gaps
✓ Identifies missing prerequisites
✓ Estimates learner proficiency
✓ Adapts explanations dynamically
Every student gets a different learning experience
#EdTech #MachineLearning #PersonalizedLearning