Keystone Applied Intelligence

Governed AI infrastructure for regulated industries

Authorization-first retrieval, governed conversational agents, and portable evaluation infrastructure — for environments where wrong answers, unauthorized retrieval, or unaudited actions create legal, regulatory, operational, or safety risk.

Runs on customer-controlled infrastructure with local models. No external API dependency for core operation. Air-gap capable.

Demo access demo.getkeystone.ai · operator1 / demo123 · try: “What atmospheric testing is required before entering a confined space?”
01 Platform

A platform, not a point solution

Keystone is one platform spanning three workloads for regulated AI systems. They are not separate products bolted together — they share a common governance substrate, so identity, task state, coordination, and proof behave the same way across all three.

Keystone Engage Conversational

Governed conversational agents for regulated customer interaction

Severity-tier escalation, per-step evidence gating, and human-in-the-loop routing for high-consequence interactions — enforced in the orchestration layer, not the prompt.

Public eval: keystone-engage/agent-v1 · 100/100 (70/70 core-regression, 25/25 architecture, 5/5 edge-case)

Public repo: keystone-engage

Keystone Counsel Retrieval

Authorization-first retrieval for legal, financial, and compliance content

Classification-aware, client-isolated access control enforced by a WHERE clause in the retrieval query, evidence-backed answers with citations, and fail-closed refusal when evidence is insufficient.

Evidence: cross-client isolation regression test (denial verified on both filtered and unfiltered paths). No sealed eval baseline yet.

Public repo: keystone-counsel

Keystone Verify Evaluation

Portable evaluation infrastructure for governed AI systems

Structured, reproducible eval artifacts with a public failing-to-passing lineage — the evidence behind every capability claim, reusable across workloads and deployments. The ledger seals and versions them.

Public ledger: keystone-ledger

Shared substrate

One governance substrate underneath all three

Every workload inherits the same control-plane primitives, so governance is consistent by construction rather than re-implemented per product.

  • Agent identity and roles
  • Task lifecycle state management
  • Event-driven coordination
  • Dispatch abstraction
  • Reproducible evaluation artifacts, sealed in the ledger
02 Why this exists

Operational discipline for the LLM era

The enterprise contact-center industry solved severity-tier escalation, per-step validation, compliance logging, confidence-threshold refusal, and policy-aware routing long before LLM products rediscovered these needs. Keystone carries that discipline into the LLM era and implements it as architecture rather than prompt instructions.

Contact centers solved
What LLM products rediscovered
Severity-tier escalation
Human-in-the-loop for high-consequence actions
Per-step validation
Evidence gating between agent steps
Compliance logging
Integrity-checked audit trails
Confidence-threshold refusal
Fail-closed when evidence is insufficient
Policy-aware routing
Authorization-first retrieval and dispatch
03 Controls

Structural controls, not prompt instructions

Governance is enforced in the retrieval flow, the orchestration layer, the database model, and the audit chain — outside the language model. A control that lives only in a prompt can be reworded away. These cannot.

Evidence-backed answers

Every answer cites the specific source document and section it draws from. No unsupported assertions.

Retrieval flow

Fail-closed refusal

When evidence is insufficient, or a query is outside corpus scope, the system refuses rather than guessing — with separate, audited refusal reasons.

Retrieval flow

Query-time role-based access control

Roles are checked before retrieval runs. In the governed-retrieval extensions, unauthorized documents are excluded by a WHERE clause in the database query and never enter the query context; the conversational extension enforces corpus-scope authorization. Both fail closed.

Database model

Human review for high-consequence actions

Severity-tier routing sends high-consequence interactions to a human before they complete, not after.

Orchestration

Factual consistency scoring

A dedicated model (Vectara HHEM-2.1-Open) independently scores whether each generated answer is supported by retrieved evidence. Below-threshold answers are withheld. When the scorer model is unavailable, the score is recorded as unavailable and the answer is not blocked on it.

Orchestration

Audit trail with integrity checks

Every query and every refusal is written to an append-only audit trail. Engage and Counsel hash-chain entries with SHA-256 and verify the full chain on replay. Gov signs each entry with a keyed HMAC-SHA256 integrity check (per-record verification, not a chain walk).

Audit chain
04 Proof

Claims backed by public eval artifacts

Capability claims map to versioned evaluation artifacts in the public ledger — quantitative baselines, a documented remediation history, and a failing-to-passing lineage. Proof artifacts, not roadmap language.

keystone-core/retrieval-v1 · 2026-04-11
  • Precision@1 0.75 · MRR 0.79
  • Corpus: 53 documents, 2,674 chunks (Alberta OHS)
  • Hybrid retrieval: pgvector + full-text search
  • Adversarial ACL: 8/8 blocked, 0 leaks
  • Fail-closed: 5/6 (83%)
  • Audit receipts logged (hash-chain fields empty pending re-seal)
keystone-core/agent-v1 · 2026-05-20
  • Governed agent extension
  • Corpus: 135 documents, 23,684 chunks
  • 186 eval cases, 12 categories
  • 558 executions, 0 failures
  • 153 strict pass · 33 characterization
  • 4 system bugs found by eval — all fixed & re-verified
Remediation · failing-to-passing lineage FC-005 domain scope guard, demo-grade, merged 2026-05-17. A pre-retrieval gate that refuses out-of-corpus queries (emissions regulations, workers comp, federal tax, IT procurement). It targets the retrieval-v1 FC-005 failure, where a greenhouse-gas query returned Part 36 mine-gas chunks via embedding overlap. A re-verified passing run is not yet sealed. The full taxonomy-based two-stage domain gate is not yet implemented; this is a partial, documented remediation. The prior failing run (keystone-core/agent-v0) is published alongside the passing one. See the eval ledger.
Not claimed
  • Enterprise HA or disaster recovery
  • Multi-node or distributed deployment
  • OIDC/SAML production identity integration
  • Third-party penetration testing
  • WCAG accessibility compliance

Recent work: governed-incident-agent — generative UI for governed agent actions, built at AI Tinkerers Global Hackathon, May 2026.

05 About

Built by Arnaldo Sepulveda

Keystone Applied Intelligence is the independent engineering practice of Arnaldo Sepulveda. The approach comes directly from nearly 13 years at Genesys, building the AI intelligence layer of enterprise contact centers — environments that required production reliability, security controls, and the ability to prove what happened, when, and who authorized it.

That work spanned Knowledge Center retrieval, the chat suite, e-services, and contact intelligence, across regulated and public-sector deployments. The governance patterns Keystone enforces are not new inventions — they are hard-won operational discipline, carried into the LLM era and implemented as architecture.

  • Practice Independent engineering practice of Arnaldo Sepulveda · based in Canada
  • Background ~13 years at Genesys — AI intelligence layer of enterprise contact centers
  • Scope Knowledge Center retrieval · chat suite · e-services · contact intelligence
  • Domains Regulated and public-sector deployment experience
  • More arnaldosepulveda.com