Sumit Kumar

Hey builders, I’m Sumit.
I’m building Tokenshark. Think of it as Wireshark for local LLMs.
It’s a zero-cloud-GPU visual debugger and inspector that lets engineers step through transformer/hybrid architectures, diagnose 4-bit quantization loss (KL divergence / rounding errors), inspect LoRA diffs, and stress-test model guardrails via activation steering. It also ships with interactive lessons for devs moving from API-wrappers to local model deployment.

I’m currently running live technical workshops and demos on local LLM internals, quantization diagnostics, and model red-teaming for engineering teams and tech meetups.

If you're deploying open-weight models on-prem, organizing an AI meetup in Singapore, or just want to geek out on model internals, drop a comment or DM me!