I built this around a question I keep coming back to: what should an agent be allowed to change on its own? SentinelOps reads alerts, logs, and commits, prepares a patch, tests it in a sandbox, and stops for review when the blast radius is unclear.

I wanted a morning briefing without giving an AI permission to “clean up” my inbox. MailDM uses read-only Gmail OAuth, summarizes messages for each user’s timezone, and sends the result through Discord—without sending, deleting, archiving, or marking mail as read.

This started as an attempt to avoid tying an agent to one model vendor. The repository separates provider, runtime, host, storage, secrets, and scheduler, and includes a tested FastAPI service with streaming, Redis state, tenant rate limits, API-key auth, and Prometheus metrics.

My open-source work is mostly practical maintenance: fixing the edge case that breaks a provider, making forked CI behave, or turning an idea into a runnable MCP app. Recent contributions include Cognee embedding support, Headroom tool detection, and NitroStack’s Zerogravity sample.

I built DexProbe to make repetitive mobile-security checks easier to run and compare. It brings together APK inspection, Flutter unpacking, decompilation, evidence collection, and scoring into a pipeline I can inspect and extend.

This is the record of my seven months in the bi0s security club at Amrita. The notes cover the things I was actively learning: GDB, x86-64 assembly, Linux challenges, APK analysis, CTFs, and the slow process of learning to read a binary.