i write C++ because i like knowing where my memory lives. python and js pay the bills and ship the demos. rust is the new toy on my desk — half-learned, half-loved, fully humbling.

right now i'm deep in mechanistic interpretability — poking at transformer internals, residual streams, attention heads, trying to figure out what these models are actually doing under the hood instead of what they say they're doing. there's a difference, and the gap is where most of the interesting stuff hides.

on the side i'm picking up CUDA, getting comfortable with kernels, memory hierarchies, and the general vibe of writing code that runs on a few thousand threads at once. parallel programming on GPU clusters is genuinely fun in a way i didn't expect — it's like puzzle-solving where the puzzle pushes back.

i've built LLMs from scratch (yes, the whole stack — tokenizer, attention, training loop, the works) and shipped a few AI-native apps. i like building things that don't work yet more than things that already do.

// the stack, roughly
auto me = Developer {
    .heart       = "C++",
    .daily       = {"Python", "JavaScript"},
    .learning    = "Rust",
    .tinkering   = {"mech interp", "CUDA", "GPU clusters"},
    .built       = {"LLMs from scratch", "AI-native apps"},
    .obsession   = "what's actually happening inside the model"
};

when i think about LLMs, i don't think about chatbots. i think about oracles.

the ancient ones — Delphi, Cumae, the seers kings consulted before marching to war. people brought them the biggest questions of their lives and walked away convinced. and for centuries, empires were built and burned on the strength of that conviction.

the oracles weren't always wrong. that's the unsettling part. they were often right, just often enough that doubting them felt unreasonable. and when they were wrong, the ashes were enormous.

LLMs feel like that to me. they're genuinely intelligent — not in the cute "look what it did" way, but in the this-is-actually-reasoning way. and we're starting to consult them on real things. real decisions. real lives.

but here's what keeps me up:

• evaluation isn't a solved problem. it's not even close. the benchmarks shift, the failure modes shift, the goalposts shift. we're grading the oracle with a ruler that keeps changing length.
• there's no moral compass in there. not a broken one — none. whatever ethics show up are stitched on from the outside, and the stitching is loose.
• and yet they're smart enough to be persuasive. which is the exact combination that got empires into trouble the first time around.

so that's the thread running through most of what i build — interpretability work, evaluation harnesses, sycophancy probes, CoT faithfulness tests. not because i think AI is bad, but because i think we don't actually know what it's doing yet, and pretending we do is how the ashes happen.

• Instructional Assistant — DSA & Discrete Math · Penn State · Aug 2024 – present
• AI Research Intern — Network IDS · Anna University CEG · Feb 2023 – Jul 2024
• Platform Engineer Intern · Tvastr · AWS compliance automation across 1000+ accounts
• ML Engineer · Lokesh Agro Foods · image classification for produce grading, family-owned organic foods business

languages

C++ for serious work · Python and JS for everything else · Rust because it's time

ML / LLM

infra