Job Application Tracker Full-stack, multi-user job-application tracker with a Kanban pipeline board, stats dashboard, and deadline/follow-up reminders.

• Problem solved: manually tracking applications across spreadsheets doesn't scale — this gives a single pipeline view with reminders so nothing falls through.
• Tech: React 19, TypeScript, Vite, Tailwind CSS, FastAPI, PostgreSQL (SQLAlchemy 2.0 + Alembic), JWT auth
• Key features: drag-and-drop Kanban (dnd-kit), Recharts dashboard, CI running pytest against a real Postgres instance, deployed frontend (Vercel) + API (Render)
• 🔗 Live app · Repo

Customer Churn Prediction Predicts telecom customer churn (IBM Telco dataset, ~7,043 customers) with a served model, SHAP explanations, and a SQL/BI analytics layer.

• Problem solved: flagging at-risk customers before they churn, optimized for recall on the churn class rather than misleading raw accuracy.
• Tech: Python, scikit-learn, XGBoost, SHAP, FastAPI, SQL, Streamlit
• Key features: Logistic Regression vs Random Forest vs XGBoost compared head-to-head (XGBoost selected, ROC-AUC 0.8405), SMOTE applied correctly (train-only), live SHAP top-factors per prediction, a separate SQL+Streamlit BI dashboard for ops teams
• 🔗 Repo

Credit Card Fraud Detection End-to-end fraud classifier on a severely imbalanced dataset (284,807 transactions, ~0.17% fraud), with a Streamlit scoring app.

• Problem solved: a naive model hits 99.8% "accuracy" by never predicting fraud — this is tuned and evaluated specifically for extreme class imbalance.
• Tech: Python, scikit-learn, XGBoost, imbalanced-learn (SMOTE), Streamlit
• Key features: SMOTE vs class-weighting compared across 3 models (6 combinations), best model (XGBoost, class-weighted) hits F1 0.865 / Recall 0.847 / ROC-AUC 0.965, adjustable decision threshold exposed in the app since the right cutoff is a business tradeoff, not a fixed 0.5
• 🔗 Repo

PathForge — PDF Q&A (RAG) Upload a PDF, ask questions in a chat interface, get answers grounded in the document with cited page numbers.

• Problem solved: LLMs hallucinate on domain documents — this grounds every answer in retrieved, citable source pages.
• Tech: Next.js, FastAPI, Pinecone (vector storage), Gemini (embeddings + chat)
• Key features: split frontend/backend deployment architecture (stateless Vercel frontend, persistent Render backend) chosen deliberately for RAG's stateful vector-store requirement
• 🔗 Live demo · Repo

Document Forgery Detection (CV) CPU-only pipeline fine-tuning ResNet18 to classify image regions as authentic or tampered (copy-move / splicing).

• Problem solved: manipulated ID/document photos are a common KYC and onboarding fraud vector — this flags copy-move and splicing forgeries using forensic cues.
• Tech: PyTorch, ResNet18 (transfer learning), Error Level Analysis (ELA) preprocessing, PIL/NumPy
• Key features: fully synthetic, procedurally-generated training data — no external dataset downloads required, runs top-to-bottom on a laptop CPU in minutes
• 🔗 Repo

Email Job Scheduler Production-style batch email scheduler: thousands of scheduled emails, per-sender spacing, hourly caps, and a live queue dashboard.

• Problem solved: sending bulk scheduled email reliably, resilient to restarts, without blowing through provider rate limits.
• Tech: Express + BullMQ (TypeScript), Prisma, Redis, Elasticsearch, Next.js dashboard, Docker Compose
• Key features: API and worker run as separate, independently-scalable processes; live searchable queue dashboard; Slack alerts on rate-limit; Google OAuth login
• 🔗 Repo