Where I test AI systems instead of just using them: what they can do, how to check them, and where human judgment has to stay in charge. Sentinel is public now. Krinel has a working prototype. Two more projects are on the bench.
Agent security lab, open source under Apache-2.0
Does the action match the authorization?
An open-source testnet lab. Before an AI agent’s proposed blockchain transaction runs, Sentinel checks it against the mandate a person actually signed, and only that exact, checked call can execute.
Testnet lab; no production security guarantee.
A transaction can call an allowed contract and stay under a spending limit while sending the purchase to the wrong person. Sentinel gives that failure a reproducible test: does this proposed action conform to what the owner signed?
The proposer has no signing authority. An evaluator checks the call’s decoded arguments and simulated effects against the mandate. A signer in a separate process attests the decision, and an execution contract checks that the submitted call matches the signed record.
You can follow that sequence through the source code, run the failure cases, and verify signed decision records. The repository includes the tests, adversarial findings and corrections behind its claims, so you can examine both the mechanism and the evidence for it.
Human authorization. Exact calls. Inspectable evidence.
Local AI research prototype, in development
Small models. Bounded judgment.
I’m building Krinel to give ordinary software a small, local judgment layer. It recommends from a fixed set of options while the application keeps control of permissions, approvals, and execution.
The working CPU prototype needs no cloud service or API key. Its synthetic demo shows recommendations, recorded actions, replay, and fallback when the model is unavailable.
Text-based judgment is still under evaluation. Public code and a runnable demo have not been released.
Neither has a public demo yet.
In early build
Build a small language model to understand how it works and where it fails. Then explore a local, retrieval-assisted workflow with each step open to inspection.
In design
An AI-assisted mathematics research process combining language models, symbolic tools, search, and formal verification. Progress must leave a checkable trail from conjecture to proof.
Sentinel is available to explore now. More project code and write-ups will be linked here as they’re released.