I build and probe machine-learning systems — from mechanistic interpretability of transformers to adaptive algorithms and production AI. Second-year CS @ VIT Vellore.
A CS student who treats models as objects to understand, not just train — drawn to the intersection of AI/ML, systems, and AI safety.
I'm Shane, a B.Tech Computer Science student at VIT Vellore (2024–2028). My work spans probing the internals of large transformers, building runtime-adaptive algorithms in C++, and deploying ML on real financial data during my internship at South Indian Bank.
I care about interpretability and alignment — recently completing BlueDot Impact's Future of AI course — and I like shipping things that are measurable, fast, and honest about what they do.
Probed DeepMind's 270M-parameter Searchless Chess transformer for 8 human chess concepts — material balance, king safety, check detection — across all 17 layers, revealing the model front-loads interpretable representations in early layers before encoding action-values later.
A runtime system that analyses graph structure and dynamically selects the optimal shortest-path algorithm (BFS, Dijkstra, Dial's). A Random Forest classifier predicts the fastest algorithm with 88% accuracy, cutting execution time dramatically on sparse graphs.
A peer-to-peer travel coordination portal for students, with mandatory university-email verification and secure PNR logging — streamlining holiday travel logistics through safe, verified communication.
View on GitHub →AI safety, alignment, and the societal implications of advanced AI.
Verified certificate in core web technologies and development principles.