I'm Chirag - I build things closer to the metal than most coursework asks for: a matching engine that clears orders by price and time, a shell that talks to the OS directly, models trained to explain their own mistakes.
I read Computer Science at IIT Gandhinagar, after starting out in Mechanical Engineering and deciding partway through that I'd rather spend my time inside compilers and order books than CAD files.
These days most of what I build sits at the boundary between systems programming and applied ML: how a matching engine keeps a book consistent under concurrent updates, how a shell turns a line of text into a running process, how a CNN's mistakes trace back to a specific architectural choice. I like the kind of problem where you can't hand-wave the details — the code either handles the edge case or it doesn't.
Outside of coursework, I run student body elections for IIT Gandhinagar's 3,000-plus students as Election Commissioner — which turns out to be its own kind of systems problem, just with people instead of packets.
Two systems projects in C++, two applied ML projects in Python. Built end to end, not templated off a tutorial.
A matching engine that mirrors real exchange mechanics: incoming orders are matched against resting liquidity by price-time priority, with correct handling of partial fills and live book state across multiple price levels.
Price levels are keyed on std::map, giving O(log n) access for both inserting new orders and querying book depth at scale.
A working shell built on raw syscalls — fork, execvp, pipe, dup2, waitpid — supporting multi-stage piping (cmd1 | cmd2), I/O redirection (>, <), and background execution (&).
The point was never to reimplement bash — it was to stop treating process creation as a black box.
Trained a CNN and an MLP on the same 60,000-image, 10-class CIFAR-10 split and found a 31-point accuracy gap (84% vs. 53%) — then went looking for why, rather than stopping at the number.
Ablation studies plus visualised feature maps and loss curves traced the gap to specific architectural choices, and informed a dropout-regularisation fix.
A set of controlled experiments comparing Batch GD, SGD, and Mini-batch GD on convergence and stability, plus a Decision Tree (info-gain, pruning), KNN, Logistic Regression, and SVM evaluated with k-fold cross-validation.
PCA was used for dimensionality reduction and 2D visualisation, turning model comparisons into something that could actually be read off a plot.
Spearheading end-to-end Student Body Elections for 3,000+ students — nominations, campaign regulations, and result declaration — while coordinating a team of election officers.
Organised flagship events at IIT Gandhinagar's annual cultural festival, managing logistics, vendors, and on-ground execution for 500+ attendees.