casestudy Production Agentic Operations © Operating an Agentic OS: Unattended AI You Can Trust A fleet of Claude agents works these systems every night, unattended — and the autonomy ladder tops out below auto-merge on purpose, because the merge decision stays human. Four bounded loops, each with a contract and a named kill switch; 21 process lessons logged, every one priced by the failure that taught it. Read the full one-page case study → Agentic AIAI OperationsClaudeAI SafetyPrompt InjectionCost Governance View details →
github Patient-Safety Gated · Human Sign-Off © PACCA: Healthcare Prior Authorization AI Platform More than one in four physicians (26%) report that prior authorization has led to a serious adverse event for a patient in their care, and the HHS Inspector General found 13% of sampled Medicare Advantage denials were for services that actually met coverage rules. PACCA answers with a five-agent clinical platform whose seven deterministic pre-flight gates and three governance gates override model confidence on experimental treatments, rare conditions, and conflicting guidelines — every ambiguous case routed to a Medical Director. Pre-production; synthetic cases only. PythonFastAPIClaude APIChromaDBReactDocker View details →
github 10–14 hrs → 90–120 min per evaluation © Adult Learning Coaching Agent (ALCA) Cuts a coaching evaluation cycle from 10–14 hours to 90–120 minutes, and ties every observation to a timestamped transcript citation instead of a black-box score. A multi-agent pipeline on Claude Sonnet 4.5 evaluates pacing, engagement, content structure, and adult-learning principle application — the same rubric on every session, rather than quality drifting by evaluator. Baseline-versus-target figures; synthetic instructor data only. PythonClaude APIAssemblyAIAgentic AIEdTechFastAPI View details →
whitepaper 784 tests · 105 cases · 3 gates © SOTA Agentic PRD: PACCA Healthcare AI Prior authorization decides whether a patient gets treated. PACCA automates that decision with five AI agents, then refuses to trust them: three deterministic gates — a per-run intent contract, a minimum-necessary scope guard, and an evidence-grounding detector — stand between the model and any automated approval. Any gate trips, a human decides. 105 synthetic cases. Every limitation written down. Agentic AIAI GovernanceHealthcare AIMulti-Agent SystemsClinical ValidationHIPAA / SaMD View details →
whitepaper 7 Artifacts · 9-Phase Lifecycle © CRISP-AG: Enterprise Agentic AI Governance Framework Closes the gap between what ISO/IEC 42001, ISO/IEC 42005, and NIST AI RMF require and what teams must actually produce to deploy agentic AI safely: seven governance artifacts, a nine-phase lifecycle, and a five-class delegation taxonomy — now with a Responsible-AI approver and a consequential-decision flag. v3.0 adds agent-identity, ISO/IEC 42005 impact-assessment, and workforce-impact artifacts, and specifies mitigations for cascading-agent-failure (ASI08) and human-trust-exploitation (ASI09) threats. Presented as design propositions and implementation guidance — explicitly not yet empirically validated. AI GovernanceAgentic AIISO 42001ISO 42005NIST AI RMFMulti-Agent View details →
whitepaper 175 REQ-IDs · Audited Against the Build © Specification-Driven Design: PACCA Healthcare AI (v3.0) Denial is now a human-only terminal state — derived, not preferred. CMS-0057-F, 42 CFR §422.101(c), and state AI statutes bar an algorithm from being the sole basis of an adverse determination, so agents may approve or escalate and never deny. v3.0 also retires the temperature-determinism contracts — sampling parameters are being withdrawn across the model line — and makes confidence routing inoperative without a current calibration record. 175 requirement IDs, coverage computed rather than asserted: half carry a verification test, and the other half are named individually. The specification was then run against the implementation as an audit instrument, and the evaluation harness turned out to be measuring 20 of the 42 cases its own gate scored — a gap that had been quietly flattering every variance number the repository quoted. Agentic AIAI GovernanceHealthcare AICMS-0057-FHIPAAPrompt Injection View details →
guidance 5 Modules · Operations Curriculum © Agentic Ops: Running AI Agents You Can Trust (Mini Course) The answer to the interview question most candidates cannot handle: what stops it from doing something destructive when nobody is watching? Five modules on rules that check themselves, bounded loops with named kill switches, verifiers that hold no tools and so cannot be talked into anything, autonomy earned from logged evidence, and cost-aware model routing. Agentic AI Mastery teaches you to build agents; this teaches you to operate the ones you cannot watch. Agentic AIAI OperationsCourseClaudeAI SafetyLoop Engineering View details →
whitepaper 7 Capability Pillars · Primary-Source Cited © The Production Harness: Engineering AI Agents You Can Walk Away From Anyone can prompt an agent. The engineering is everything around the model call. This white paper documents the seven-pillar discipline behind a production Claude Code harness — context layering priced by what it costs every call, an Evidence Ledger that ends false "done," hooks that turn advisory rules into guarantees, and unattended loops whose autonomy tops out below auto-merge by design. Every practice traced to Anthropic's own engineering guidance; every pillar earned through a documented failure. Read the full white paper → Agentic AIHarness EngineeringClaude CodeAI SafetyVerificationContext Engineering View details →
github 1.5–2 hrs saved per lesson © K-12 Lesson Planning Assistant (LPA) Returns 1.5–2 hours per lesson to teachers while keeping pedagogical authorship with them: a 4-Pass workflow where the teacher reviews the structure, customizes for their students, adds their own teaching style, then polishes the materials. Differentiation scaffolds for struggling learners, advanced learners, and ELL students generate with every plan. A prototype — its quality and adoption figures are stated targets, not measured results. PythonFastAPIClaude APIK-12 EducationSQLiteEdTech View details →
whitepaper Six Teacher Competencies © Agentic AI Implementation Guide for K-12 Education Translates corporate AI-management competency into the classroom: six teacher competencies, from pedagogical context assembly through recognizing where the capability frontier ends. FERPA-tiered data handling and equity-first design — proactive bias mitigation, multilingual support, within-district equity monitoring — are treated as requirements rather than afterthoughts. K-12 EducationAI TrainingFERPAEquityProfessional Development View details →
patent US 6,850,988 • 2005 Dynamic E-Commerce Click Stream Analysis US Patent 6,850,988: System for interpreting navigation patterns to optimize e-commerce strategies. E-CommerceUser BehaviorAnalyticsPersonalization View details →
patent US 6,839,229 Large-Grained Database Concurrency Management US Patent 6,839,229: Log monitor with dynamically re-definable business logic for data sharing. DatabaseConcurrencyData WarehousesMemory Management View details →