RateLockr
Distributed API Rate-Limiting Gateway
Node.js, Express, TypeScript, Redis, Docker, React
GitHub · Live demo
- Built an API gateway with two rate-limiting strategies, token bucket and sliding window log, in TypeScript and Redis. Under k6 load at 100 concurrent users it sustained about 290 requests per second and blocked every over-limit call across 3,000 requests.
- Moved counter updates into Redis Lua scripts so two gateway instances cannot double-count the same request. Replaced a wildcard SCAN with a fixed 30-bucket time series, which cut Redis round trips per request.
- Added a fail-open path so traffic keeps flowing when Redis is unreachable, and streamed live counters to a React and Recharts dashboard through a browser event bus.
Trashium
Waste Collection and Recycling Platform
Next.js, TypeScript, Supabase, Tailwind, Leaflet
GitHub · Live site · Sole developer on a 5-person team
- Built the full application in Next.js and TypeScript with three user roles, household, crew and admin, on Supabase Postgres. Access rules run as Row-Level Security policies in the database instead of only in the UI.
- Shipped the crew view with a live Leaflet map, zone-based pickup queues and geo-tagged proof photos served through short-lived signed URLs, plus an admin analytics dashboard in Recharts.
- Ran the team's pricing regression inside the booking request instead of hosting a separate inference service, backed by a Postgres fallback table so a quote never fails outright.
- Added English, Hindi and Bengali language support and made the app an installable PWA that works offline.
- Owned testing for the platform: Playwright end-to-end suites, Vitest unit tests, and alpha and beta rounds before each release.
Foggy Lane Detection
Computer Vision Pipeline
Python, OpenCV, NumPy, Streamlit
GitHub · Live demo
- Wrote a Python and OpenCV pipeline that removes fog from road images using dark channel prior, atmospheric light estimation and bilateral filtering, then detects lanes with Canny edge detection and the Hough transform.
- On low-visibility test clips the dehazed frames recovered lane markings that the raw frames missed entirely.
- Deployed it on Streamlit with automatic image downscaling, which roughly halved processing time per frame.