Data infrastructure, decision analytics, and technical PM delivery across pipelines, dashboards, cloud workspaces, and applied ML systems.
I work where technical execution meets delivery: raw messy data, validation checks, status visibility, cross-functional handoffs, and decision tools people can use. Past work includes regulated banking delivery, road-safety ETL on Databricks, Kubernetes research workspaces, explainable risk classification, public-health surveillance, and recent decision-analytics case studies.
The proof is strongest when each story shows what was broken, what I built, and what changed after it shipped.
Different teams care about different signals. The same body of work maps cleanly into technical PM, data infrastructure, decision analytics, applied ML, fintech risk, and public-sector analytics roles.
I am strongest where the job needs both technical execution and delivery discipline: messy inputs, reliable pipelines, validation evidence, stakeholder updates, handoff notes, dashboards, and cross-functional coordination.
Each card starts with the problem, then the system built, the result, and why that work maps to real hiring needs.
Python · FastAPI · DuckDB · LangChain · LSTM · React · CDC SODA API
Public-health analysts had millions of CDC surveillance records but no fast way to query trends, forecast demand, or get answers without waiting for a data pull.
FastAPI backend ingesting live CDC SODA data into DuckDB, LSTM hospitalization-surge forecasting, and a LangChain RAG interface for plain-English questions.
Real-time surveillance plus accessible querying is exactly the category health systems, insurers, and public agencies are trying to operationalize.
Python · SQL · React · notebooks · pytest · product analytics
A wearable-membership team needs to know whether guided onboarding should become the default experience without overstating early wins as long-term retention proof.
Fictional, synthetic A/B-style case study with KPI contracts, SQL models, executed notebook, dashboard, requirements, validation report, and testable decision rules.
Guided onboarding showed 89 more activated members per 1,000 signups, while day-60 paid retention stayed uncertain. Recommendation: staged rollout with a preserved comparison group.
Shows business analysis, product analytics, metric definitions, guardrails, privacy-aware scope, and executive communication in one reviewable package.
Python · Streamlit · Statcast · pybaseball · pytest · validation
Postgame pitcher-fatigue analysis can mistake a noisy velocity slope for a real performance decline if pitch type, game context, coverage, and survivor bias are ignored.
Statcast analysis package and Streamlit dashboard using game-by-pitch-type baselines, complete-game holdouts, bootstrap uncertainty, coverage warnings, and reproducible notebooks.
Ran the complete pipeline on Logan Webb's 2024 season: 33 eligible starts, no supported 1 mph threshold, 0.75 mph held-out MAE, 23 passing tests, and 75% branch-aware coverage.
Designed as decision support for historical review, not injury diagnosis or an automatic pitching-change model.
Python · scenario modeling · public sources · executive brief · CSV outputs
Cloud infrastructure planning needs to compare markets across demand, power feasibility, cooling exposure, fiber, permitting, and risk without pretending public data is internal strategy.
Auditable market-screening model for AI and cloud datacenter capacity planning, with base, power-first, and inference-first scenarios plus explicit assumptions and caveats.
Base case ranks Northern Virginia, Atlanta, and Hillsboro/Portland highest, then shows how rankings shift when power delivery or inference-fit weights change.
Recruiter signal for market intelligence, infrastructure strategy, AI capacity planning, assumption discipline, and concise stakeholder readouts.
Python · SQL · Random Forest · bootstrap uncertainty · stakeholder brief
Basketball decisions need more than a leaderboard: lineup, role, tactical context, sample size, defensive risk, and uncertainty all shape whether a recommendation is usable.
Synthetic possession-level work sample with SQL summaries, random forest possession-value modeling, time-based validation, player role matrix, and lineup decision board.
Generated 4,428 synthetic possessions across 12 lineup profiles, with validation MAE of 0.269 points per possession and recommendations framed through uncertainty bands.
Shows how to turn a vague sports-operations question into a model, validation plan, SQL outputs, and a coaching-friendly decision artifact.
FinBERT · BiLSTM · XGBoost · SHAP · LIME · Streamlit
Banks use static repayment rules that miss distress signals in customer communication, miss payment anomalies, and cannot explain decisions cleanly to auditors.
Three-branch ML system: FinBERT for communication distress scoring, BiLSTM for payment sequence anomalies, and XGBoost for five-tier recovery classification.
93.4% accuracy, 0.990 AUC-ROC, 26.1% SHAP signal across 30,000 customers, with explainability and demographic fairness review built in.
Group capstone. My positioning emphasizes the ML pipeline, explainability, risk-classification logic, and analyst-facing delivery.
Python · Flask · SQLite · scikit-learn · scipy · Monte Carlo · VaR/CVaR
Individual investors and small teams have limited access to institutional-grade risk analytics like VaR, stress testing, scenario analysis, and sentiment overlays.
Full-stack platform with Monte Carlo simulation, Sharpe/Sortino ratios, historical stress-test scenarios, NLP news sentiment, and reporting through a Flask API.
Direct signal for quant risk, wealth management, banking analytics, and financial modeling roles.
Apache Spark · Spark SQL · MLlib · Databricks · Python
Campus transit systems schedule around average demand, then over-provision on normal days and under-provision for events.
Predictive demand model on 700MB+ simulated transit data using PySpark and Databricks MLlib, with route clustering, peak-load analysis, and sustainability metrics.
85% decision-tree accuracy on ridership prediction, with architecture that maps to transit, logistics, and operations scheduling teams.
NodeMCU · PMS5003 · MQ-7 · LSTM · XGBoost · Python
Air-quality monitoring in underserved areas depends on expensive fixed stations, leaving residents without real-time hyper-local insight into what they are breathing.
IoT pipeline from NodeMCU sensors through LSTM/XGBoost forecasting to a monitoring dashboard, with weekly refresh cadence and drift checks.
92%+ forecasting accuracy against official pollution-board readings. Same sensor-to-model-to-dashboard pattern used in industrial IoT, smart city, and environmental monitoring work.
Banking, research, and public-sector work with quantified delivery across pipelines, ML evaluation, dashboards, and cross-functional execution.
Grouped by the work recruiters screen for: SQL, machine learning, data engineering, cloud systems, dashboards, and stakeholder delivery.
Python, SQL, PySpark, Databricks, Spark SQL, DuckDB, pandas, ETL/ELT, data quality checks, pipeline optimization.
PyTorch, Hugging Face, scikit-learn, XGBoost, LSTM, Transformers, SHAP, LIME, forecasting, model evaluation.
FastAPI, Flask, Streamlit, React, Plotly, Power BI, Tableau, REST APIs, stakeholder-facing dashboards.
Kubernetes, Azure Databricks, reproducible cloud workspaces, API-backed analytics systems, Git workflows.
KPI contracts, experiment interpretation, scenario modeling, uncertainty, financial risk, anomaly detection, Monte Carlo simulation, VaR/CVaR.
Stakeholder communication, project lifecycle, BRDs, UAT, issue tracking, status updates, release notes, handoff documentation, dashboard storytelling.
Graduate training in data science with a computer science foundation from NIT Delhi.
M.P.S. Data Science | GPA: 3.7/4.0 | Full-time STEM graduate program | May 2026
B.Tech Computer Science and Engineering | GPA: 3.0/4.0 | May 2023
Boston and remote. STEM OPT authorized. Available immediately. Useful for teams that need data moved, tested, explained, coordinated, and shipped into decision workflows people can use.
