Shaurya Kesarwani

Final-year CS undergrad, Bengaluru. I ship complete, production-minded systems — low-latency backends, explainable ML, real-time full-stack products, and tools that make messy Indian public data usable. Everything below is tested, documented, and reproducible with one command.

Selected work

LinkForge

Go · PostgreSQL · Redis · k6 · Prometheus
~26,500 req/s · p99 ≈ 5.4 ms · 99.99% cache-hit · zero errors

High-throughput URL shortener: cache-aside Redis, rate limiting, async click analytics, full observability. Load-verified on a single instance.

DevBoard

React/TypeScript · Express · Prisma · Socket.IO
Live multi-user sync · JWT refresh rotation with reuse detection

Real-time collaborative Kanban: optimistic drag-and-drop with rollback, owner/editor/viewer roles enforced on every endpoint, live activity feed.

Credit Risk Explainer

Python · XGBoost · SHAP · FastAPI · Streamlit
ROC-AUC 0.871 · PR-AUC 0.404 vs 0.067 baseline · Brier 0.049

Loan-default prediction where every score ships with a SHAP explanation — because a score you can't explain is a score you can't audit.

Hidden Mutual-Fund Fees

Python · pandas · statsmodels
1,084 fund pairs · Direct wins 99% · ≈₹27 lakh lost per 20-yr SIP

Measured the hidden distributor commission in Indian mutual funds from official NAV data. The gap between identical Regular/Direct plans is the fee.

Wafer Yield Predictor

Python · XGBoost · imbalanced-learn · SHAP
86% FAIL recall at a 6.6% failure rate (accuracy is the wrong metric)

Fab pass/fail prediction from ~590 sensor traces, with per-wafer SHAP attribution so engineers can trace a yield excursion to its sensors.

Civic data tools (4 repos)

Python/FastAPI · React/TypeScript · Docker
133 tests · three-state engines that never overclaim

SchemeFinder (scheme eligibility) · SafeDose (spurious-drug batches) · MandiCompass (mandi net prices) · JalCheck (water contamination)

Open source