SHREYANSH KUMAR
Building intelligent systems
that scale and deliver.
Building intelligent systems
that scale and deliver.
I've always been drawn to the moment when something abstract becomes real. Being able to bring an idea that someone has been carrying around to life is what sparked my passion for AI and ML. Studying Computer Science at Bennett University gave me the foundation and Applied Data Analytics at Boston University sharpened how I think about problems. But working with data consultancies and diving into real problems is where I learned that clean models are only half the solution. If the delivery doesn't land, none of it matters.
This has shaped how I work now. At Droisys in Las Vegas, working across the full stack of an ML-powered surveillance system, my care goes beyond just the models. The craft of logging, observability, pipelines that don't break at 3am is just as important to me. A problem can have ten paths through it, but I'm most interested in the one that's efficient and worth learning from.
Outside of work, I follow F1 with the same attention I bring to a production incident, with the same care for every race, every sector time, and every strategy call. When I'm not watching races I'm usually deep in a film or a playlist, drawn to the kind of storytelling that trusts silence as much as noise.
Hardened observability across a multi-module Spring Boot codebase with SLF4J structured logging. Contributed to the Table Guard Web API exposing JWT-secured REST resources for casino surveillance. Supported ETL + rule-engine pipelines ingesting and transforming table game data for real-time risk signals.
Designed an AI-powered multi-agent tutoring platform that improved learner outcomes by 30% and lifted engagement by 35%. Architected a scalable RAG pipeline with recursive chunking and metadata-aware vector search, cutting topic confusion by 42%. Built intelligent document processing with 98% extraction accuracy and 90% reduction in API costs.
Led development of a Python-based time series forecasting tool using ARIMA, SARIMA, SARIMAX, VARMAX, and RNN models, reducing exploratory time by 70%. Automated hyperparameter tuning across 50+ iterations per dataset, cutting delivery from 3 weeks to 4 days. Enhanced model accuracy by 12% with iterative feature engineering.
Predicted sales using ARIMA and Prophet models, reducing overstock by 12% and stockouts by 9%. Improved product discovery with semantic embedding search, boosting session duration by 30%. Developed K-Means/DBSCAN customer segments for personalized campaigns that raised repeat purchase rates by 18%.
Large-scale retrieval-augmented recommendation engine over the 8GB+ Yelp dataset. Semantic retrieval paired with LLM-based generation delivers context-aware dining suggestions across 10K+ businesses with low-latency querying.
End-to-end ML classification pipeline with advanced feature engineering and hyperparameter tuning. SMOTE for class imbalance and Boruta for feature selection raised predictive performance by ~18%. Dimensionality reduction cut training time by ~30% without accuracy loss.
Cinema is where I go to think differently. I'm drawn to films that treat silence as seriously as dialogue — where the frame itself is doing the storytelling.
Music is the background layer running under everything. The right album changes how a problem feels — I think in playlists as much as pipelines.
F1 is the only sport where engineering and instinct collide at 300km/h. I follow every race, every team radio, every sector time — and there's only one driver worth talking about.
Sketching and photography are how I slow down. I'm drawn to texture, contrast, and the quiet details most people walk past — the same instincts that make a good dataset interesting.
Florals
in Ink
Pen &
Paper
NEW YORK
BOSTON