Student by enrollment, researcher by curiosity, occasional casualty of exam season

Saurab Mishra

I started ML for fun and ended up deep in research papers, failed runs, and exam week chaos.

Explore my work

System status: stable until exams and reproducibility checks begin.

Saurab Mishra standing under a towering tree.

Research mode: confidently confused

Professional research posture. Unprofessional internal monologue.

My interests want LLMs, systems, and building cool things. My course list wants abstract math proofs, surprise exam patterns, and emotional damage. Every semester feels like two different operating systems fighting for the same RAM, and I am the unlucky process manager asking: why is this theorem in my AI timeline right before finals?

"Yes, this experiment was definitely planned and not discovered at 2:13 AM."

87% Chance the code worked once and now owns the roadmap.

3 AM Peak model insight, peak confusion, same timestamp.

Infinite Confidence level while saying "small quick fix".

Step 1

Input Embedding

Every great model starts with raw, unoptimized noise. In my case: an Indian kid who mistakenly thought 'playing video games' naturally translated to 'enjoying calculus.' I somehow ended up at IISER Thiruvanathapuram studying topics that work better as sedatives at parties. The tokenization phase was brutal, mostly breaking down my remaining innocence into highly dimensional academic trauma.

Saurab = DropOut(social_skills) + PosEncode(IISER_TVM)
batch_size = 1 · focus_level = overflow_error

Raw Human Input

IISER Thiruvanathapuram

Curiosity

Coffee

Research Student

Step 2

Multi-Head Obsession

Like a proper attention mechanism, I try to attend to everything simultaneously: LLMs, multi-agent systems, XAI, and why my code worked yesterday but not today. I have multiple heads, but absolutely zero attention span for anything outside an IDE. My Query is "how do I force this model to stop lying?", my Key is chronic sleep deprivation, and my Value... well, my value is still converging.

Obsession(Q, K, V) = softmax(Papers × ArXiv^T / √sleep)
heads = [LLM, MAS, NLP, XAI] · social_life = nan

ATTENTION SIMULATION LIVE

1 Project input → Q, K, V x·W_q

2 Compute Q·K^T scores [0.8, 0.2, 0.6, 0.9]

3 Scale by √d_k ÷ 8.0

4 Apply softmax [0.31, 0.12, 0.22, 0.35]

5 Multiply attention × V context ✓

Research Student

LLMs

NLP

XAI

ArXiv at 3 AM

"I Swear I'll Sleep"

Step 3

Reality Normalization

Ah, Reality Normalization, that incredibly humbling layer where your 'groundbreaking' idea gets absolutely dismantled by Reviewer 2. Residual connections? That's just me falling back on the excuse "at least the Python script runs" when the mathematical proofs collapse. The skip connection to the campus coffee machine is literally the only thing preventing a vanishing gradient of my will to live.

sanity = LayerNorm(ego - Reviewer2(scathing_remarks))
ε = 1e-5 · skip_to(coffee) preserves gradient

Sleep-Deprived

Deadlines

Revision Loop ↩

Somehow Functional

Step 4

The Build Phase

This is where the alleged "thinking" takes place. While normal people expand their social circles, I exclusively expand my dimension of hidden vectors. I build multi-agent systems, dissect LLMs, and maintain a GitHub profile that clearly screams "I have no offline hobbies." PyTorch is my primary love language; TensorFlow is that toxic ex we agreed never to text again.

Projects(x) = GELU(ambition × PyTorch + bias)
hidden_talents = 3072 → completely unmarketable

Motivated

CrewAI + LLaMA 3.3

RAG Pipelines

"It Works on My Machine"

Step 5

Output Layer

After all that expensive compute, what is the output? A research student clutching IBM certifications, a Credly profile with actual badges, and the concerning social habit of explaining neural architectures to captive audiences who definitely did not ask. I'm actively working on LLM interpretability, because apparently, accepting that AI works by magic isn't scientifically rigorous enough. The forward pass is incomplete. Loss is still nonzero. Send help.

status = still_training(actively, please_wait)
P(touching_grass) = 1e-9

All That Work

IBM Certified

Research Student

Hire Me → Maybe

"Curiosity Is All You Need"

Naïve Ambition

IISER Thiruvanathapuram

+ Curiosity Boost

Multi-Agent Obsession

Reality Check & Norm

LLM / RAG Projects

Burnout Prevention

Explainability Pipeline

Research Student Output

∞×

The Architecture That Ruined My Sleep

I subjected myself to this architecture in 2024, and here I am in 2026 still waiting for the code to compile. The diagram on the left illustrates the tragic pipeline that took a healthy, functioning human and trapped them in a latent space of multi-agent systems and Large Language Models. I study exactly what goes on inside these cursed blocks, specifically, why my scripts run flawlessly at 3 AM but spontaneously combust by sunrise.

Because if we're going to give neural networks the ability to confidently gaslight humanity, the least we can do is figure out how they're doing it. And that's my problem now.

"Curiosity Is All You Need"
S. Mishra et al., IISER TVM 2026

Saurab Mishra

Professional research posture. Unprofessional internal monologue.

The Forward Pass of Saurab Mishra

Input Embedding

Multi-Head Obsession

Reality Normalization

The Build Phase

Output Layer

The Architecture That Ruined My Sleep

Explore

Work