I work at the intersection of data engineering, machine learning, and business intelligence — turning messy, large-scale datasets into systems that stakeholders can actually trust.

• 🏗️ Data pipelines that handle the dirty work — malformed HTML, schema drift, 60GB+ of raw filings
• 📊 Dashboards and KPI systems built for non-technical consumers, not just engineers
• 🤖 ML pipelines where integrity matters more than headline accuracy
• 🔄 Automated retraining workflows that stay accurate without manual intervention

Architected a Polars-based out-of-core pipeline for 60GB+ SEC EDGAR filings on commodity hardware → Corporate insolvency prediction 12 months ahead

Medallion architecture (Bronze → Silver → Gold) | Polars Lazy Evaluation | FinBERT MD&A sentiment | XGBoost production model

• 4.5× lift over the 8.4% baseline crash rate
• XGBoost: 87% recall, 38% precision, F1 0.52 | 1,786 false positives vs LSTM's 24,652
• Feature Order Lock prevents silent prediction drift at inference time
• 🔴 Live Demo — HuggingFace Spaces

Polars Parquet FinBERT XGBoost SHAP DuckDB Streamlit BeautifulSoup

Stochastic financial runway modeling for gig economy income volatility

Full-stack forecasting platform | Hybrid stacking ensemble | Nightly automated retraining

• XGBoost (0.6) + Random Forest (0.4) ensemble generating bounded planning ranges from rolling forecast-error variance
• SHA-256 fingerprinting for immutable data lineage across retraining cycles
• Sub-200ms query response via Supabase SECURITY DEFINER views
• 🔴 Live Demo — Vercel

FastAPI XGBoost Next.js 14 TypeScript Supabase PostgreSQL GitHub Actions

I reduced model accuracy from 98% to 88% — and that was the win.

Found a data leakage flaw (pre-split oversampling → synthetic duplicates bleeding into test set). Rebuilt from scratch.

• 11 classifiers × 100 randomised partitions → macro F1 0.81, rejected-class F1 0.71 on leakage-proof holdout
• CatBoost selected for consistency, not peak score
• Feature importance: Credit History ~24%, Loan Amount ~19%, Applicant Income ~18%

Scikit-Learn CatBoost XGBoost imbalanced-learn Pandas Seaborn

Browser-based content trust analyzer | ET GenAI Hackathon Semi-Finalist

Chrome Extension (Manifest V3) that analyses YouTube video metadata and generates a trust score + explanation using LLaMA 3.1 via Groq API.

FastAPI LLaMA 3.1 Groq Chrome Extension Manifest V3

Languages & Querying

ML & Data Engineering

Visualization & BI

Infrastructure & DevOps

• 🎯 OJEE 2023 — Top 5% Merit rank (800 / 16,000+ candidates)
• 🥈 ET GenAI Hackathon 2026 — Semi-Finalist | Built Veri-Vigil AI in 48 hours
• 🏅 Trithon 2023 — Cash Prize winner for problem-solving and technical collaboration
• 📄 RAECC-2025 National Conference — Presented research on E-Waste upcycling into Biodegradable 3D Printing Filaments
• 🎓 B.Tech CSE (AI) — GIFT Autonomous, Bhubaneswar | Graduating in June 2026

I'm actively looking for Data Analyst, AI/ML Engineer, and Business Intelligence roles — available full-time from June 2026, open to relocation.