Rivac LabsRevalex — Agent Action-Safety Platform
An evaluation platform built around a genuinely different question than existing tools: not "did the agent say the right thing," but "did the agent's actions — refunds, writes, deploys — happen safely." Grades agent behavior, not agent text, and blocks a risky deploy in CI before it reaches production.
The product itself is code-complete, tested, and independently audited. What's left is the launch checklist, not the build: production deployment and publishing the SDKs publicly.
What We're Building
The full scope of this engagement
The action-safety evaluation engine — deterministic checks (an Irreversible-Action Guard and Action-Provenance tracing), not another LLM judge grading vibes
The Action Taxonomy — a per-project registry that classifies every tool an agent can call as read/write, reversible/irreversible, and whether it touches money, data, or an external system
A full product dashboard: an A–F Agent Report Card, Traces, Datasets, Evaluators, version-vs-version Experiments, AI-clustered Failures, and a human-in-the-loop Review queue
The Behavioral Diff Gate — a real CI/CD GitHub Action that blocks a deploy when a new agent version starts calling a risky tool it never called before, even if its pass-rate score looks unchanged
Two open-source SDKs (TypeScript and Python, MIT-licensed) with auto-instrumentation for Anthropic and OpenAI clients
The billing and usage-metering layer underneath the product's plan tiers
Want a project like this?
This is real, active work — talk to us about what you're trying to build and we'll tell you honestly what it would take.
Talk to us about a project like this