AI infrastructure for biosecurity
Hey folks! đź‘‹
We’re building Mohi, a debugging assistant for AI agents. If you've ever spent hours scrolling through 500+ LLM calls asking “what the hell happened?” — this is for you.
The Problem
Agents are complicated — tools, memory, web search, nested loops, and hallucinations. And when something breaks, it’s like untangling spaghetti. You’re stuck scrolling through flat logs, guessing which of the 35 steps silently failed.
It's slow, brittle, and makes you second-guess your whole stack.
The Fix
Mohi makes debugging agents fast, and actually usable.
Just tag key steps in your agent with our lightweight SDK (think: tool calls, memory access, context switches, decision logics), and we build a clean visual graph of what happened under the hood. Then we go a step further and show you:
- where your agent last succeeded
- where it went off the rails (hallucinated, failed tool, wrong branch)
- and how to re-run broken steps with new inputs
It’s not just observability. It's a diagnosis.
Who We Are
Previously, Hao worked on AI and distributed systems observability at AWS, and Evan worked on automated testing at Google and AI evaluations at SecureBio. We met at MIT and worked on a bunch of projects together, eventually building agents ourselves and seeing firsthand how painful debugging can be — so we built Mohi to fix it.
Our Ask
If you’re building complex agents — we’d love to chat!
