undergrad at IIIT Bhopal. i'm into systems, performance engineering, and ml.
i like knowing what's actually happening under the hood. if something works, i wanna know why. if it's slow, i wanna see where the time goes (profiler, not vibes).
right now:
- writing a lot of C/C++ and getting more comfortable with low-level stuff
- poking around linux, kafka, computer networks, and monitoring
- writing CUDA kernels, willingly, ngl
- reading about computer architecture, distributed systems, and how frontier ai really work
- ml systems stuff: llm post-training, RLHF, distributed training
- competitive programming on the side (LeetCode Guardian, CodeChef 4★, Codeforces Specialist)
| project | what it is |
|---|---|
| overveckill | a vector database in C++, built from scratch to understand what's going on inside one |
| cracked.c | neural nets and autodiff from scratch in C. no framework, no safety net |
| cuda-reranker | custom CUDA attention kernel + profiling. biggest lesson so far: a faster kernel ≠ a faster pipeline |
| repo-trace | corrective RAG for navigating unfamiliar codebases. retrieve, verify, retry when the answer looks off |
| flash-attention | digging into attention implementations and the GPU memory traffic problem |
| train-maxx | learning LLM training and how to scale it past a single GPU |
| llm-post-training | post-training algorithms implemented from scratch, following Deep-ML's RL fine-tuning for LLMs path. less magic, more math |
Attention Is Not What You Think (private, for now)
i'm trying to look at attention in transformers as something you can actually test instead of a black box. controlled experiments, real hypotheses, and being honest about what the results do and don't say.
still early. i'll share more when there's something real to show.
- understand the internals before leaning on the abstraction
- profile first. a guess isn't a benchmark
- read the source, read the paper, try to reproduce the result
- when something breaks, that's usually where the learning is
me: this should take 20 min
6 hours later:
- read a paper
- stared at assembly
- derived something on paper
- questioned a few life choices
- got a 3% speedup, called it a win
- languages: C, C++, Python, SQL, Bash
- systems: Linux, Git, Docker, PostgreSQL, Redis
- performance / GPU: CUDA, PyTorch profiling, kernel optimization
- ml: Transformers, RAG, RL, LLM post-training
- learning next: computer architecture, distributed training, Kafka, monitoring, RLHF, and the math behind transformers
chess ♟️ · lifting 🏋️ · running 🏃 · math 📐
if you're into systems, performance, ml research, or building things from scratch, feel free to reach out. always down to nerd out.