ProjectsThings I've built.

Multi-tenant HR SaaS where every company gets its own subdomain and an isolated Postgres schema, with wildcard TLS provisioned automatically through Caddy. The AI layer carries it: NLP resume scoring against job descriptions, live WebSocket voice interviews that generate scored reports, one-line-prompt JD generation, and per-tenant RAG policy chatbots on ChromaDB. Async top to bottom on FastAPI, Celery, and Redis, billed through Razorpay and deployed in a single Docker Compose command.

FastAPINext.jsTypeScriptTailwindCSSPythonPostgreSQLRedisChromaDBGemini 2.5 FlashDockerCaddyGoogle CloudRazorpay

Two-sided services marketplace where customers book verified professionals and admins vet every provider before they go live. JWT-gated dashboards for the three roles, Razorpay checkout, and PDF invoices generated on the fly with WeasyPrint. Celery and Redis run the async and scheduled jobs so booking, notifications, and reporting never block a request.

PythonFlaskVue.jsJavaScriptBootstrapChart.jsRedisCeleryRazorpay

Ad marketplace where sponsors spin up campaigns and both sides negotiate ad requests back and forth until a deal lands. Runs on a RESTful Flask API with bcrypt-hashed auth, three roles, and validated sign-ups. Chart.js dashboards give sponsors and influencers live visibility into campaigns, requests, and spend.

PythonFlaskJinjaHTMLCSSJavaScriptBootstrapChart.jsSQLite

Two-layer LightGBM voting ensemble that predicts Windows malware infection from antivirus telemetry, ranked #1 of 1,700+ in the IIT Madras ML competition. The edge came from feature engineering, not model complexity: across 76 features, an engineered temporal signal (the gap between antivirus and OS install) became a top-5 predictor. Nine baselines benchmarked, eight tuned LightGBM variants stacked into a hard-voting meta-classifier, landing 79% malware recall.

PythonNumPyPandasMatplotlibSeabornscikit-learnXGBoostLightGBMJupyter