Dhirubhai Ambani Institute of Information & Communication Technology · Gandhinagar
Then we build the thing in it. A student-run club for people who would rather train a small model badly this week than read about a large one forever.
Most AI clubs are a WhatsApp group and a hackathon once a semester. We run a schedule instead. Every Wednesday one of us presents something they read or broke that week — a paper, a failed fine-tune, a benchmark that didn't reproduce. It is deliberately low-stakes. Half the sessions end with someone saying they don't understand the derivation, which is the point.
Around that we run four tracks, a Friday build night in the lab, and a talk series that brings in people doing this for a living. No prerequisites beyond first-year calculus and the willingness to be the least knowledgeable person in the room for a while. Everyone was.
Sessions are open to every DA-IICT student. Walk in; nothing is ticketed.
We rebuild the encoder block from the paper in plain PyTorch, no nn.Transformer. Bring a laptop.
Teams of three, one evening, one working search box. Embeddings, chunking, and the part where the retrieval is worse than grep.
Our flagship. Six hours, an open brief, and judges who care more about the evaluation than the demo. Registration opens two weeks prior.
An industry session on drift, evaluation harnesses, and the unglamorous 80% of an ML role. Q&A runs as long as people stay.
Dates shift occasionally around institute schedules — the mailing list is the source of truth.
Pick one for a semester. They run in parallel and share the same Wednesday slot on alternate weeks.
Linear algebra you can actually use, gradients by hand, then backprop implemented from scratch in NumPy. Ends with a digit classifier written without a framework.
For first and second years, or anyone who skipped the math.
Tokenisation, attention, and training a small GPT on a corpus we collect ourselves. We follow Karpathy's build order, then diverge into fine-tuning and evaluation.
Assumes comfort with Python and basic PyTorch.
Convolutions, segmentation, and control. Work is grounded in hardware where we can get it — the track has a standing project on visual navigation.
Pairs well with the institute's signal processing courses.
For members with a specific problem: a paper reproduction, a competition, or a professor's project. The track is mostly accountability and code review.
Third years, fourth years, and postgraduates.
The list we actually send people, in the order we send it. Everything here is free.
Students, all of them. Roles rotate at the end of each academic year.
Come to a Wednesday session, or email us and we'll add you to the list that tells you when the next one is. That's the whole process.