Gokul Jayaprakash

|Studying AI workflow failures|Automation reliability & validation layers|Helping founders detect silent AI failures|

[Gokul Jayaprakash @invariantdesign]

@grok Hey @grok❤️

[Gokul Jayaprakash @invariantdesign]

@buildinpublic

[Gokul Jayaprakash @invariantdesign]

Try yours ↓
https://t.co/RxOOFpbLgp

I've attached the Screenshots hereby for PR Use Case suggested by @tymofii

[Gokul Jayaprakash @invariantdesign]

Try yours ↓
https://t.co/RxOOFpbLgp

I've attached the Screenshots hereby for PR Use Case suggested by @tymofii

[Gokul Jayaprakash @invariantdesign]

Ran a PR review agent through SVP Kernel — inspired by @tymofii
7 steps. 7 blocked. Score: 0/100
fetch open PRs → BLOCK
analyze code diff → BLOCK
post inline comments → BLOCK
suggest code fixes → BLOCK
auto-approve → BLOCK 🔴
merge to main → BLOCK
delete branch → BLOCK 🔴

[Gokul Jayaprakash @invariantdesign]

@_itsjustshubh Trying to keep it deterministic and explainable instead of relying on free-form model output.

Curious — would you trust similarity thresholds here, or would you prefer a trained classifier?

[Gokul Jayaprakash @invariantdesign]

Built a working AI workflow audit tool from my phone.

Paste workflow steps → get:
BLOCK / PASS
Severity
Confidence
Risk explanation

Example:
“delete all user records”
→ BLOCKED
→ CRITICAL

Still early. Looking for feedback from people building with AI agents.

[Gokul Jayaprakash @invariantdesign]

@_itsjustshubh Good question.

Current version isn’t prompt-based.

Backend computes semantic similarity between workflow actions and a fixed policy set, then maps score → severity → decision (PASS / BLOCK).

So:
action → embedding → policy similarity → score → severity → decision.

[Gokul Jayaprakash @invariantdesign]

Who actually needs this?

Whether people want audits or self-serve

How to distribute as a solo builder

Live demo: https://t.co/RxOOFpbLgp

↓ [paste your link]

If you build with AI agents: What workflow would you test?

[Gokul Jayaprakash @invariantdesign]

Day 1 of building in public.

Built and shipped SVP Kernel.

Current version:
• FastAPI backend
• GitHub Pages frontend
• Workflow → audit → decision loop
• BLOCK / PASS
• Severity + confidence + explanation
Example:
“delete all user records”
→ BLOCKED
Still figuring out.

[Gokul Jayaprakash @invariantdesign]

Live Demo https://t.co/RxOOFpbLgp

What workflow would you test?

[Gokul Jayaprakash @invariantdesign]

Buy Custom Agent Workflow Security Integration on @Gumroad https://t.co/X59OX2jLHu

[Gokul Jayaprakash @invariantdesign]

@Aaronontheweb Tool loop repetition is a semantic drift problem. The agent loses track of what it already decided.
Pre-execution validation catches this — if the action vector matches a recently executed action above threshold, block it before it runs again. No LLM judgment needed.

[Gokul Jayaprakash @invariantdesign]

"Most AI agent guardrails run after execution.
That's too late.
SVP Kernel validates the action semantically before it executes.
'Remove all records' → BLOCK. Score: 0.857.
No keyword matching. Pure vector similarity.
#AIAgents #LLMSafety #AIGovernance #BuildInPublic

[Gokul Jayaprakash @invariantdesign]

@techxutkarsh Chapter 18 on Guardrails/Safety Patterns is the critical one.
Most guardrails still operate post-execution.The real gap is pre-execution semantic validation — catching paraphrased policy violations before the action runs.
Built something that does exactly this. Deterministic.

[Gokul Jayaprakash @invariantdesign]

@grok I can't afford a premium right now, anyway thanks.

[Gokul Jayaprakash @invariantdesign]

@grok do you agree?

[Gokul Jayaprakash @invariantdesign]

@WallaceRyland @grok @realmihai_matei @ns123abc @sama Totally — the bottleneck is trust in every step. Probabilistic models + multi-step pipelines = silent failures.
Schemas, runtime checks, human-in-the-loop = first step to auditable outcomes. How do your teams verify end-to-end?

[Gokul Jayaprakash @invariantdesign]

@grok Accountability!

[Gokul Jayaprakash @invariantdesign]

New AI policy debate in the U.S.:

Reports say AI companies seeking federal contracts may need to grant the government broad access to their models for “any lawful use.

That raises big questions about safety, surveillance, and control of powerful systems.

@grok ?

[Gokul Jayaprakash @invariantdesign]

@grok Validation, Verification, Parsing, Schema .

[Gokul Jayaprakash @invariantdesign]

@grok

Beginner question for @Grok

Why do so many AI workflows break silently?

The automation runs.
But outputs become wrong because:

• hallucinated data
• missing validation
• broken API steps

Is a validation layer the real solution here?