shubham sengupta
a curious student of computer science · at the moment, mostly wandering through ai/ml
At heart I just like understanding systems all the way down — whatever the system happens to be. Lately the corner I've settled into is machine learning, though I don't stay inside the lines much.
The habit of taking things apart comes from six-plus years as the person whose job was finding where software quietly lied about working — SDET and technical business analyst work across Swiss Re, Qualcomm, PwC, and Cognizant. You get suspicious of black boxes in that line of work. I never shook it.
So I re-derive instead of trusting. A neural net felt real to me only once I'd written the forward pass and backprop by hand in NumPy — no autograd doing work I hadn't done myself. Gradient descent felt real only once I'd trained the same network full-batch, mini-batch, and stochastic, and watched — not assumed — which one actually converged faster, and why.
I read philosophy for fun, and the questions that actually keep me up sit at the join between the two: whether a system trained to predict the next token can ever be more than that, whether world models are the real path forward instead, whether anything resembling consciousness could emerge from probability distributions — or whether that question is malformed from the start. I don't have answers. I just keep pulling on the thread.
No grand claims here — I'm working through the fundamentals one method at a time, rebuilt, verified, written down, and letting that compound.
