I'm a Data Analyst and ML Engineer based in Bengaluru, working at the intersection of SQL-driven business intelligence and applied machine learning. I started in mechanical engineering — a B.E. from GCET and two internships analyzing production data on factory floors at Elecon Engineering and Jinal Engineering, where I traced 12–15% excess cycle time back to specific process inefficiencies. That operational grounding is still how I approach data: find the number that's actually wrong before reaching for a model.

I've completed a PGP in Data Science with a GenAI specialization at Great Lakes Executive Learning. I build full end-to-end projects rather than notebooks that stop at a metric — data cleaning through deployment, with a documented decision trail at every step.

Latest completed project: NeuroTrade, a full-stack ML platform predicting Buy/Sell/Hold signals for 20 US stocks, deployed live at neurotrade-signals.vercel.app.

Open to: Data Analyst · ML Engineer · BI Analyst · Data Scientist roles — Bengaluru or Remote.

Languages

Machine Learning & Data Science

Generative AI

Visualization & BI

Web & Deployment

Databases & Tools

Repository · Live App

Python XGBoost LightGBM LSTM FastAPI Next.js Groq / LLaMA 3.3 70B

Full-stack ML platform predicting Buy/Sell/Hold signals for 20 US stocks, trained on 5 years of OHLCV data (29,100 rows) with 17 engineered technical indicators (RSI, MACD, Bollinger Bands, ATR, volume features).

• Trained 3 base models — XGBoost, LightGBM, and a 20-day-sequence LSTM — combined via a Logistic Regression stacking ensemble on out-of-fold predictions
• Enforced strict no-leakage methodology: 5-day purge gap at CV fold boundaries, date-based TimeSeriesSplit on unique trading dates, and per-ticker LSTM sequencing
• Held out 2024 as a true validation set, opened exactly once for honest evaluation — XGBoost delivered the best result at F1 Macro 0.3739, ahead of the 0.33 random baseline
• Deployed end-to-end: FastAPI backend (Hugging Face Spaces) serving a Next.js/TypeScript frontend (Vercel), with a Groq LLaMA 3.3 70B assistant generating plain-language signal explanations
• Co-built with a second ML engineer, splitting modeling and deployment responsibilities

Repository

Python XGBoost Scikit-Learn Pandas Tableau Public

Ball-by-ball win probability predictor for IPL 2nd innings, trained on 278,000+ deliveries across 1,169 matches (2008–2025), visualized through an interactive 4-view Tableau dashboard.

• Ran a 7-step data cleaning pipeline (dropped super overs, standardized 5 team rebrands and 30+ venue name variants)
• Engineered 10 features, the most important being Pressure Index (Required Run Rate − Current Run Rate), which alone drove 52% of the model's decisions
• Compared Logistic Regression, Random Forest, and XGBoost — selected XGBoost not for highest accuracy, but for the lowest False Negatives (3,765 vs 4,291 for LR), since predicting a team is losing when they're actually winning is the costlier broadcast error
• Exported predictions to a 4-view Tableau Public dashboard: Win Probability Curve, Team Performance, Pressure Index Analysis, Match Summary

Repository · Live App

Python XGBoost SHAP Streamlit Groq / LLaMA 3.3 70B Plotly

Binary classification system trained on 433,074 Americans from the CDC BRFSS 2023 survey, identifying HIGH mortality risk from self-reported behavioral and demographic data alone — no lab tests required.

• Engineered a proxy target (high_risk) from 8 clinically validated chronic conditions, since BRFSS has no death column
• Engineered 4 domain-driven features (comorbidity_count, age_risk_tier, health_burden, ses_score), validated against published clinical literature on mortality risk
• Applied a 7-strategy logical imputation framework (peer-group medians, clinical rules, mode-only for race) across data up to 43% missing
• Tuned XGBoost (RandomizedSearchCV, 50 iterations × 3 folds) achieved ROC-AUC 0.8256 and 85.12% Recall at threshold 0.40, outperforming Logistic Regression, Random Forest, LightGBM, and a Neural Network
• Deployed as a full-stack Streamlit app with SHA-256 auth, SHAP explainability, risk gauge/radar charts, and a Groq LLaMA 3.3 70B clinical assistant

Repository

Python MySQL SQL Tableau Faker

Full-stack SaaS analytics pipeline on a synthetic 677,179-row, 5-table relational dataset, covering 29 SQL queries (KPIs, cohort retention, RFM segmentation, anomaly detection) and 5 interactive Tableau dashboards.

• Designed a normalized MySQL schema with DECIMAL precision for monetary fields, BIGINT keys for the 581K-row events table, and a composite index for join performance
• Found Enterprise CLV is 17.6× higher than Starter, and Trial churn sits at 40.5% — the highest-risk segment
• Detected a payment failure anomaly in August 2023 (31.51% failure rate, Z-score 4.51) using pure-SQL statistical detection
• Quantified a 30-day pre-churn "silence window": churned users reduce activity by 62–73% in the month before cancelling — a concrete early-warning signal for Customer Success
• Built a 12-month cohort retention matrix and 7-segment RFM model (NTILE-based) identifying ₹5,96,631 in at-risk MRR

• PGP in Data Science with Specialization in GenAI — Great Lakes Executive Learning (2024–2026) Machine Learning · Deep Learning · NLP · Generative AI & LLMs · Statistical Modeling
• B.E. Mechanical Engineering — GCET, Anand, Gujarat (2021–2024)
• Diploma in Mechanical Engineering — N.G. Patel Polytechnic, Surat (2018–2021)

Let's Connect

Open to Data Analyst, ML Engineer & BI Analyst roles — Bengaluru or Remote.