About Keystone Applied Intelligence

Keystone Applied Intelligence is an independent AI engineering and R&D practice focused on enterprise AI systems where authorization, evidence, evaluation, failure handling, and operational state matter alongside model capability.

The public engineering work includes separately composed implementations for governed retrieval, conversational-agent workflows, authorization-aware retrieval, evaluation, and runtime-governance research. These repositories contain related runtime mechanisms, but they should not be interpreted as one fully composed production runtime.

The engineering approach is influenced by more than a decade of work with enterprise contact-center systems, where routing, escalation, access control, observability, incident handling, and controlled change are operational requirements rather than optional features. Keystone applies those engineering habits to LLM-based systems while treating the limits of the current implementations explicitly.

Examples include query-time authorization before retrieval, fail-closed behavior under selected conditions, severity-based human routing in the served Engage path, implementation-specific audit mechanisms, retained evaluation artifacts, and explicit task-state mechanisms. These capabilities are not uniform across all workloads, and their presence in one implementation should not be read as evidence that every Keystone component provides the same control.

See the contact-center heritage for the engineering patterns that motivated parts of the platform design.

How to navigate these docs

Start with Architecture for the engineering architecture and the substrate model for the separate research abstraction.

Extensions covers the current public capabilities:

• keystone-engage: governed conversational-agent engineering, with a served single-agent path and an optional experimental multi-agent path.
• keystone-counsel: authorization-first retrieval with classification and client-isolation controls.
• keystone-verify: a standalone evaluation harness for compatible HTTP endpoints.

Evaluation documents the published internal baselines and the retained passing and failing artifacts available through keystone-ledger. These results are evidence about the evaluated implementations at their recorded commits; they are not independent validation or proof of production suitability.

Design explains some of the engineering lineage behind the work. Access describes which implementation and evidence surfaces are public and which deployment and infrastructure details remain private.

Runtime-governance research is developed separately under Governed Execution. Its bounded public reference implementation is Runtime Validity (Track A).

Where to go next

For the employer-facing engineering narrative and a walkthrough of the current platform work, see getkeystone.ai/platform/.

Keystone is independent engineering and research work. For technical review or hiring conversations, see arnaldosepulveda.com or LinkedIn.