About
Lelu AI
We are building authorization infrastructure for autonomous AI agents — the security layer that decides what an agent is allowed to do before it acts.
Company Overview
Company
Lelu AI
Founded
2026
Founder
Abenezer Getachew
Team
3 members + 1 advisor
Product
Authorization Engine for AI Agents
Status
Live — 2,000+ SDK downloads
SDKs
TypeScript, Python
License
Open Source (MIT)
Mission
AI agents are powerful enough to take real actions — processing refunds, modifying databases, sending emails, making business decisions. But traditional authorization systems were not built for AI. They assume deterministic, human-driven actions.
AI agents operate with uncertainty, make probabilistic decisions, and can be manipulated through prompt injection attacks. Companies need a security layer designed specifically for autonomous AI.
Lelu provides the security infrastructure that makes AI agents safe for production — allowing companies to move fast without breaking things.
What We Built
Confidence-Aware Gating
Every agent action carries a confidence score. Low-confidence actions route to human review automatically — no code change needed.
Human-in-the-Loop Review
Agents pause, queue their action, and resume once a human approves or denies. Full audit trail included.
Policy Enforcement
Write YAML or Rego policies to express complex rules. Allow, deny, or require review — per actor, action, or resource.
Complete Audit Trail
Every decision is logged with full context, input hash, output hash, latency, and confidence score for compliance and debugging.
Research
A Confidence You Can Measure
Validating Confidence-Gated Authorization for AI Agents
Authorization systems for AI agents increasingly propose gating actions on model confidence — allow a tool call when the model is confident, escalate or deny when it isn't. We test that premise directly: a reproducible AUROC benchmark over 321,105 real answers from 15 chat models shows confidence separates correct from incorrect actions with an overall AUROC of 0.709, reaching 0.83 on the strongest models (GPT-4o).
Team
Software Developer · AI Researcher
Software developer and AI researcher with deep expertise in security infrastructure, distributed systems, and autonomous agent design.
Built Lelu from the ground up — engine, SDKs, and platform — driven by the conviction that AI agents need purpose-built authorization infrastructure before they can be trusted in production.
Software Developer · UI/UX Designer
Full-stack developer and UI/UX designer responsible for Lelu's platform interface, design system, and developer experience.
Shapes how developers interact with Lelu — from the sandbox playground to the audit dashboard — ensuring security primitives feel approachable and intuitive.
Advisor
PhD Candidate · AI Researcher
PhD candidate and active AI researcher advising Lelu on the frontier of autonomous agent behavior, model safety, and AI decision-making. Brings academic depth to Lelu's approach to confidence scoring, behavioral anomaly detection, and agent trust frameworks.
Contact
Open Source — MIT License
Security infrastructure should be transparent and auditable. Self-host or use our managed platform.