Final-year CS undergrad in Bengaluru. I build complete, production-minded systems across the stack — a low-latency Go backend, a real-time full-stack product, explainable ML services, and data investigations that land concrete findings. Every project below is finished, documented and tested.

A production-minded URL shortener focused on correctness and operational maturity.

~26,500 req/s on the redirect hot path at p99 ≈ 5.4 ms, 99.99% cache hit rate, zero errors (single instance, Postgres + Redis).

Cache-aside Redis, rate limiting, async click analytics, Prometheus metrics, k6 load tests, testcontainers integration tests.

A real-time Kanban tool (Trello/Linear-style) with live multi-client sync.

JWT refresh-rotation with reuse detection, optimistic drag-and-drop with rollback, typed end-to-end contracts (shared Zod schemas) — docker compose up and you're running.

Socket.IO rooms, role-based sharing (owner/editor/viewer), live activity feed, React + Express + Prisma, Jest + Supertest.

An end-to-end loan-default model that explains every prediction with SHAP.

ROC-AUC 0.871, PR-AUC 0.404 vs 0.067 baseline, Brier 0.049 — with the decision threshold set by an explicit 5:1 miss-cost, not accuracy.

Reproducible pipeline → calibrated XGBoost → FastAPI service → Streamlit dashboard. A score you can't explain is a score you can't audit.

Measuring what India's "Regular" mutual-fund plans quietly cost investors.

Across 1,084 matched fund pairs, Direct beat Regular in 99% of cases — costing a typical equity SIP ~₹27 lakh over 20 years (18% of the corpus). Paired Wilcoxon p ≈ 10⁻¹⁷⁶.

A clean, assumption-free measurement from official NAV data — identical portfolio, two share classes, the gap is the commission.

Semiconductor fab pass/fail prediction from ~590 sensor measurements, where only 6.6% of wafers fail.

Catches 86% of defective wafers — while a naive "all-pass" model scores 93% accuracy and catches zero. Accuracy is deliberately not the headline.

Leakage-safe preprocessing, imbalance-aware model selection, per-wafer SHAP attribution so an engineer can trace a yield excursion to the responsible sensors.

Hybrid retrieval-augmented generation for Indian legal questions: FAISS + BM25 + a legal knowledge graph, fused with RRF and reranked, answering with cited provisions.

R@5 0.725 and nDCG@10 0.685 on a 1,810-query held-out benchmark, against 0.559 / 0.525 for BM25 — and grounding lifts AIBE accuracy from 64.7% to 68.0%.

Legally fine-tuned bi-encoder, four-task evaluation, FastAPI + Streamlit and a React front-end, pytest suite.

Four tested, Dockerized tools that turn buried Indian public data into something usable: government-scheme eligibility, spurious-drug batch checks, mandi net-price ranking with MSP alerts, and district water-contamination lookup.

Shared design principle: three-state engines that never overclaim or over-reassure. 133 tests across the four.

76 merged pull requests across 35 repositories. Selected:

• tinyhumansai/openhuman — the cron high-frequency warning measured the gap between two near-identical instants, so it fired for every scheduled agent job; it now measures consecutive runs
• vlang/v — template directives now match only at a word boundary
• dolthub/dolt — pinned the aborted-rebase state in enginetests; reproduced & triaged a reported error-message bug
• alibaba/open-code-review — fixed the black-screen flash on lazy route transitions with React 18 startTransition
• The-PR-Agent/pr-agent — three PRs to the GitHub and GitLab integrations
• tinylibs/tinypool — filename accepts a URL instance

Stack: Go · TypeScript/React/Node · Python (scikit-learn, XGBoost, SHAP, pandas, FastAPI) · PostgreSQL · Redis · Docker · k6 · pytest/Jest

📫 kesar.shaurya@gmail.com