Hey, I'm Vinmay.
Senior engineer with 10+ years in distributed systems and backend platforms (Java, Scala, Python).
Outside work I'm digging into how small models actually work, what AI agents can really do, and where developer tooling falls short. I like building things from scratch and figuring out why they don't behave the way I expected.
- reachscan: Static analysis CLI that
shows what an AI agent or MCP server can read, write, execute, or send, and
whether an LLM can reach it. 0.47% false-positive rate.
pipx install reachscan· What I found scanning 50 MCP servers - doppel: On-device semantic scavenger hunt using EmbeddingGemma. Runs fully in the browser. Live demo
- ratchet: Building a language model from scratch in NumPy: autograd, layers, attention, transformer. No ML framework. In progress.
- I scanned 50 MCP servers to see what they can actually do:
1 in 3 can run shell commands, 8 pair credential access with network egress,
and one passes LLM input straight to
exec(). - I built "npm audit" for AI agents: Why I built reachscan and how it decides whether an LLM can actually trigger a dangerous capability.
- Your
pip installjust stole your SSH keys: How a malicious LiteLLM release used a.pthfile to exfiltrate credentials on every Python startup, and what to do about it.