A platform for running untrusted, AI-generated code safely — the sandbox an AI agent reaches for when it needs to actually execute what it writes. Exposed as an MCP tool, so agents call it directly.
Defense-in-depth isolation: every run is jailed in an ephemeral Docker container — seccomp, dropped capabilities, a read-only root, no network, and hard CPU / memory / PID / file-descriptor limits — closing eleven distinct attack vectors.
Exactly-once execution on an at-least-once queue: a Redis Streams pipeline with consumer groups, fencing tokens, and XAUTOCLAIM crash recovery, plus Postgres advisory locks for leader-elected cleanup and idempotency keys so a retried request never runs twice.
Tech Stack: Go, Redis Streams, PostgreSQL, Docker (private repo — happy to walk through the design)
Search Bar Optimization: To ensure optimal performance and reduce load on the backend, I implemented debouncing in the search bar. This minimizes the number of API calls made as the user types the query by almost 70%.
N-nested Comments: Implemented n-nested comments, where each comment and its corresponding replies are nested, making it easier for viewers to identify the replies. Used recursive components and minimized code duplication.
Live Chat: Used API Polling to simulate a live chat like environment.
A platform developed specifically for the institute’s ROTARACT club where individuals can browse through different ongoing campaigns and support a cause by donating money to the campaign.
Admin can create / edit / activate / deactivate campaigns based on requirement. Used Paytm gateway for payments; Added support for features like option to share campaigns, ask related queries.
Implemented on demand loading for various components along with shimmers for a great user experience. Followed good coding practices.
