The most severe problems in AI.

SEVzero is how I choose what to work on: the most severe unsolved problem in AI, approached with an architectural idea instead of a patch. The first was enterprise risk. Production AI failed in ways nobody could trace, so I built a lineage layer that follows a failure back to the change behind it. The next is the one the name always pointed to: SEV 0, where the AI itself is the risk. My current idea is stable agent architectures, using Sankhya and Vedantic models of mind to give agents a coherent, ethical footing. I publish the prototypes, evals, and essays as I go.

Read the writing →

Now · Current

SEV 0: stable agent architectures

The question is how to give an AI agent a stable, coherent footing: a sense of self and values that hold up under pressure. My starting point is Sankhya’s model of the mind (manas, chitta, buddhi, ahamkara, sankalpa) as a blueprint for agent architecture. The work is prototypes and evals that test whether it makes agents more stable, published as essays on Substack.

Read the writing →

Chapter 1 · 2025–26 · Complete

Enterprise risk: AI lineage

The first problem I took on: tracing a failure in production AI back to the data, prompt, model, or code change behind it. This chapter is complete and closed to new deployments.

The problem

Blind spots cause risks.

Hidden model logic and vibe coding create dangerous gaps in oversight for modern AI workflows, leading to costly mistakes and compliance violations.

Blind spots cause risks.

Hidden model logic and vibe coding create dangerous gaps in oversight for modern AI workflows, leading to costly mistakes and compliance violations.

Failures break trust.

Severe production incidents erode user and organizational trust, which is why so many AI systems die in the proof-of-concept graveyard.

Failures break trust.

Severe production incidents erode user and organizational trust, which is why so many AI systems die in the proof-of-concept graveyard.

No clear traceability.

When AI fails, it’s unclear what data or code changes caused the incident or how to prevent a repeat, so failure patterns recur.

No clear traceability.

When AI fails, it’s unclear what data or code changes caused the incident or how to prevent a repeat, so failure patterns recur.

What I built

Model Metastore with end-to-end lineage tracking

Performance drift analyzer

Governance policy enforcer

Realtime observability dashboard

About

I’m Vishy Poosala, a former engineering leader and distinguished engineer on GenAI at Meta, previously Head of Bell Labs India. After a decade working on Facebook Messenger, Privacy, Health, Safety, and GenAI, I saw one pattern again and again: safety rarely gets the priority it deserves.

If you’re working on AI lineage or drift, or on agent safety, and want to compare notes, message me on LinkedIn. I’m not offering support or deployments.