Independent AI Safety Researcher · Senior QA Engineer · FinTech reliability background

I build deterministic evidence and accountability layers for high-risk AI-agent actions before execution.

AI-agent actions should be reviewable, replayable, and evidence-backed before execution.

My work focuses on the infrastructure layer between an AI-agent proposal and a real-world effect: tool calls, code changes, infrastructure actions, financial workflows, governance actions, and other high-impact operations.

I am exploring deterministic oversight for agentic AI systems.

Core questions:

• How can high-risk AI-agent actions be evaluated before execution?
• What evidence should be required before an agent calls a tool, changes code, modifies infrastructure, or triggers a financial/governance workflow?
• How can action traces be made replayable, tamper-checkable, and useful for human reviewers?
• How can deterministic control layers complement probabilistic model evaluations?
• What should an infra-level accountability layer look like for multi-agent systems?

Pre-execution evidence gates for high-risk AI-agent actions.

AI agent proposes action -> evidence gate -> ALLOW / BLOCK / ESCALATE

Best entry point for grant, fellowship, and AI safety reviewers:
Reviewer Start Here

Verifiable intent and action-boundary audit for AI-agent/API workflows.

valid credential != valid action != valid scope != valid reversibility != valid approval

Best entry point for action-boundary and security reviewers:
Reviewer First Screen

Causality-aware QA/CI reliability substrate for reproducible failure analysis and quality decision packets.

Best entry point for reliability and open-source infrastructure reviewers:
Reviewer First Screen

The current stack is intentionally layered:

PythiaLabs -> evidence gate
ProofPath -> intent and audit boundary
CML -> causal accountability
LTP -> trace and replay protocol
LiminalQAengineer -> reliability substrate

Shared thesis:

High-risk AI-agent actions should be inspectable before execution.

Portfolio map: AI_SAFETY_PORTFOLIO.md

These projects are experimental open-source research prototypes, not production safety infrastructure yet.

They do not claim full AI alignment, complete agent safety, certified security, regulatory compliance, or universal prevention of unsafe actions.

The current contribution is narrower and more testable:

make high-risk AI-agent actions reviewable, replayable, and evidence-backed before execution.

I have 12+ years of software QA and FinTech reliability experience, including brokerage, banking, API, WebSocket, SQL, risk, reporting, test strategy, regression prioritization, and quality process design.

This background shapes my AI safety work: I treat agent oversight as an engineering reliability problem, not only as a model-behavior problem.

For reviewers, grantmakers, and collaborators:

• PythiaLabs reviewer start
• PythiaLabs portfolio map
• ProofPath reviewer first screen
• LiminalQAengineer reviewer first screen

Email: safal0645@gmail.com
Telegram: @Alexfox14
GitHub: https://github.com/safal207

I build deterministic oversight layers that gate, audit, and explain high-risk AI-agent actions before execution.