IDEA-001 · Ranking Systems Early Stage
PRISM: Position-specific Ranking with Inter-Session Memory
What if a recommender remembered not just what you liked — but what you ignored at each specific position on the page? PRISM encodes rejection signals from past sessions into position-aware context vectors.
IDEA-002 · Transfer Learning Complete
Embeddings as a Bridge: Do Neural Networks Make Classical ML Smarter?
Can a neural network teach a Random Forest something it couldn't learn on its own? Dense embeddings from a PyTorch model are handed to XGBoost and friends — with meaningful lifts on Amazon and Yelp datasets.
IDEA-003 · Fraud Detection / Ranking Complete
Cold Start Isn't a Scoring Problem — It's a Ranking Problem
Most cold start research asks how to score a new entity. This work asks where to place it in the review queue. Using uncertainty-aware strategies (LCB, Tiered), premature Top-K insertion drops from 29.6% to near zero — at negligible NDCG cost.
IDEA-004 · NLP / Search Quality Live Demo
Query Denoising: Recovering Search Intent from Misspelled Queries
When a typo doesn't just look wrong — it means something completely different. "moth ball" sends a search engine hunting for cricket gear. A denoising model corrects the query and recovers the true intent before ranking begins.
Python
PyTorch
XGBoost
Optuna
Pandas
Scikit-learn
BERT
Jupyter
NumPy
Matplotlib
Who
Shivanshu Sirohi — data scientist in the making, obsessed with ML models, ranking systems, and ideas that actually work in the real world.
Mission
Build and document research that questions assumptions — starting with recommendations and transfer learning.
Approach
Hypothesis first. Every project starts with a question worth answering — not a model worth showing off.
Currently exploring
PRISM — position-specific ranking with inter-session memory. Also digging into attention mechanisms and Bayesian hyperparameter optimisation.
Outside the lab
Car enthusiast and LEGO builder — two passions, one philosophy. The best systems, whether engines or neural nets, are elegant by design. Even in plastic bricks.
