B.Tech Computer Science & Engineering (Data Science), Manipal University Jaipur — CGPA 8.99/10, expected 2028.

I enjoy building software where reliable AI systems, intelligent applications, and data-driven decision making intersect.

AI & Data Science Intern — XYlofy AI (June 2026 – July 2026, Remote)

• Built an end-to-end ETL pipeline in Python to ingest, transform, validate, and integrate multi-year retail datasets for downstream analytics.
• Automated reusable data-processing workflows using Pandas and NumPy, reducing manual preprocessing effort across experiments.
• Built modular data-transformation and anomaly-detection components consumed by an interactive analytics platform.
• Evaluated pipeline output quality using MAE, RMSE, and MAPE to improve data accuracy and support operational decision-making.
• Maintained pipeline code using Git-based version control and collaborative pull-request workflows.

Backend / Systems · Java, Socket Programming, Concurrent Systems, CRC32, ARQ

• Custom Layer-4 protocol over UDP with a 12-byte application header (sequence numbers, CRC32 checksum, payload length, Big-Endian).
• Selective Repeat ARQ using a thread-safe ConcurrentHashMap for the in-flight packet window — 90% goodput under 20% simulated packet loss, outperforming Go-Back-N.
• Configurable network emulation proxy for stress-testing packet drop, duplication, and reordering.

→ Repository

Backend / SDE · Java, Spring Boot, Multithreading, SQL

• High-throughput batch ingestion backend processing thousands of manufacturing telemetry events/minute under concurrent load, using atomic compute() for lock-free upserts.
• 4-state idempotent upsert logic (NEW / DEDUPED / UPDATED / IGNORED) with ingress timestamps to prevent clock-skew data corruption.
• Schema validation and boundary-check layers to intercept malformed packets before persistence.

→ Repository

AI / Backend · Python, FastAPI, AsyncIO, Claude API

• 6-agent stateful LLM inference pipeline with a shared PipelineState dataclass and per-agent fallback recovery — zero HTTP 500s under simulated full agent outage.
• Behavioral Monte Carlo risk simulation (500 runs) stress-testing financial plans against user-specific spending variables.
• Pydantic-enforced JSON schema contracts across every agent boundary, eliminating downstream parsing failures.

→ Repository

AI / Full-Stack · Next.js, TypeScript, FastAPI, Gemini API

• Research-adjacent project exploring how LLMs reason over evidence: four investigation personas, an evidence board, theory comparison, and step-by-step reasoning traces.
• FastAPI backend serving prompt-engineered investigation modes to a Next.js/TypeScript frontend.

→ Repository

AI / Data · Python, Scikit-learn, Statistical Feature Engineering

• 7-dimensional feature space (Shannon Entropy, Hjorth Mobility, CoV, Skewness, Kurtosis) from keystroke timing data.
• Benchmarked Isolation Forest, LOF, and OC-SVM — LOF achieved ROC-AUC 0.8512, F1 0.7143.
• Leave-One-User-Out validation across 10 users (cross-user mean AUC 0.8031); privacy-by-design — discards key identity at the hardware level.

→ Repository

Data Science / Full-Stack · Python, Streamlit, SARIMA, Prophet, XGBoost, Scikit-learn

• End-to-end retail demand forecasting system on 4 years of transaction data: benchmarked SARIMA, Prophet, and XGBoost head-to-head, with SARIMA winning at 20.5% MAPE.
• Dual-method anomaly detection (Isolation Forest + Z-score) across 209 weeks of sales, plus KMeans/PCA-based product segmentation into four demand tiers, each mapped to a distinct inventory strategy.
• Shipped as a live, interactive 4-page Streamlit dashboard (Sales Overview, Forecast Explorer, Anomaly Report, Product Segments) — not just a notebook, an actual decision-support tool.

→ Live Demo · → Repository

Data Analyst · Python, Scikit-learn, Pandas

• End-to-end regression workflow on residential property data: cleaning, EDA, one-hot encoding, 80/20 split.
• Compared Linear Regression and Random Forest; best model (Linear Regression) reached R² 0.653 (MAE ≈ 970K, RMSE ≈ 1.32M).
• Feature importance analysis (area, bathrooms, A/C, parking) translated into pricing recommendations.

→ Repository

Data Analyst · Python, Scikit-learn, Pandas

• Predicted attrition on the IBM HR Analytics dataset (1,470 rows) using Logistic Regression, Random Forest, and Gradient Boosting.
• Best model (Gradient Boosting) reached ROC-AUC 0.805.
• Surfaced key drivers (monthly income, overtime, stock options, job involvement) as HR-facing recommendations.

→ Repository

Active contributor, currently ranked #409 globally out of 43,587 participants (top 1%).

Sample merged PRs: fixed an XSS vulnerability in skill-badge rendering, prevented concurrent-request race conditions in an AI dev assistant, and fixed EXIF orientation handling for uploaded images — contributing production-quality fixes across security, concurrency, image processing, and data correctness.

→ GSSoC Profile

• Hackathon Finalist: Offline National Finalist, IIT Bombay Upskill India Hackathon · Finalist, India Innovates 2026 (presented at Bharat Mandapam, New Delhi) · Round 2 Qualifier, EY Techathon
• Global Delegate: Selected Delegate, HPAIR 2026 (Harvard Project for Asian & International Relations)
• Certifications: NPTEL Programming in Java — Elite Gold (Top 90th percentile)
• Leadership: Joint Head of Corporate Affairs, IEEE Computer Society MUJ — secured 5+ industry sponsorships for Genesis 5.0 (20,000+ attendees)

Consistently building backend systems, AI infrastructure, and open-source software.