Home · Writing

Writing

Decision-ready writing on production AI systems, agent architecture, model evaluation, context, cloud platforms and governed delivery in financial services.

Architecture

Production AI systems design
Agent Memory as an Attack Surface: Provenance, Quarantine, DriftArchitectureAn Agentic Test Migration Factory with Repository GatesArchitectureAnatomy of a Production Multi-Agent SystemArchitectureContext Caches under Banking Privacy ConstraintsDecision-Grade Agentic SystemsContext Compilers for Policy-Filtered RetrievalDecision-Grade Agentic SystemsContext is the New ComputeArchitectureDecision Decomposition in Commercial Credit: Evidence, Arithmetic, Policy, Narrative and AuthorityArchitectureDesigning Multi-Agent Topologies with Computable Blast RadiusArchitectureDeterministic-First Agents: The Certainty Gradient for Governed Banking AutomationDecision-Grade Agentic SystemsEvent-Sourced Relationship Intelligence Without a Shadow Customer RecordArchitectureFreshness, Supersession and Valid-Time EvidenceDecision-Grade Agentic SystemsGoverned Context Architecture: Reconstruction, Provenance and Active Memory for Long-Horizon AgentsDecision-Grade Agentic SystemsGraphRAG, Relational State or Vector Search?Decision-Grade Agentic SystemsLegal Document Investigation as Reconciliation, not SummarisationArchitectureMultimodal Context for Lending and Financial CrimeDecision-Grade Agentic SystemsThree Principals, One Decision: Agent Identity and Delegated Authority in Regulated SystemsArchitecture

Deployment

Production controls, evaluation and operating evidence
Agent Reliability Atlas: 171 Production Failures, Controls and Recovery PathsDeploymentAML Alert Triage with a Hard SAR Boundary: Evidence Assembly Without Delegating SuspicionDecision-Grade Agentic SystemsCalibrated Cascade Routing in Production: Small Models First, Confident EscalationDeploymentContinuous Control Evidence: From Cloud Configuration to Auditor WorkpaperDeploymentDrift: Behavioural Sentinels and Why Tool Selection Can Lead Output MetricsDeploymentKYC Periodic Review as an Evidence-Bound Workflow: Events, Provenance and Human AuthorityDecision-Grade Agentic SystemsMeasuring Hybrid Retrieval Beyond Recall@kDecision-Grade Agentic SystemsPopulation-Scale Outcome Testing for Customer InteractionsDeploymentRCSA Evidence Assembly Without Delegating the RatingDeploymentReal-Time Fraud Investigation Agents: Fast Evidence, Bounded Action and Human AuthorityDecision-Grade Agentic SystemsRegulatory Change as a Traceable Obligation GraphDeploymentShipping a Private Skill Model Foundry: Distillation, LoRA, Calibration, and Red Teaming Inside a VPCDeploymentShipping LLM Systems in Regulated EnvironmentsDeploymentThe Eval Pipeline is the ProductDeploymentThe Evaluation Stack for Agentic Systems: Golden Cases, Adversarial Scenarios, and Process EvidenceDeployment

Google Cloud

Agentic AI on GCP
BigQuery as the Agent's Structured Memory: Medallion Pipelines, CDC and Row-Level SecurityGoogle CloudBigQuery as the Memory Layer for Agent PlatformsGoogle CloudBuilding Bank-Grade Agents on Agent Runtime: LangGraph and a Controlled Path to ADKGoogle CloudGemini Enterprise Agent Platform: What Actually Ships in a BankGoogle CloudRetrieval on GCP: Hybrid Dense and Sparse Search with RRF and Cross-Encoder RerankGoogle CloudVPC Service Controls and Workload Identity for Agentic EstatesGoogle Cloud

AWS

Agentic AI on Amazon Web Services
Agentic Architectures on AWS: AgentCore, Durable Workflows, and Where Lambda EndsAWSAmazon Bedrock AgentCore in Production: A Regulated-Industry Field ReportAWSCost Engineering Agentic Workloads on AWS: Bedrock Inference and AgentCoreAWSIdentity and Policy for Agents on AWS: AgentCore, Cedar, IAM, and Scoped AuthorityAWSReading the AWS Well-Architected Agentic AI Lens in PracticeAWS

Design

Interfaces and knowledge for human-agent teams
Designing Human-Agent Handoffs for Regulatory ScrutinyDesignDesigning Interfaces for Human-Agent TeamsDesignEvidence-First UX: Showing Claims with Their ProvenanceDesignKnowledge Architecture as a Design DisciplineDesignThe Autonomy Ledger: Making Agent Permissions LegibleDesign

Research

Current research translated into design decisions
Agent Research 2023–2026: What Held up in ProductionResearchBefore the Agent Acts: Governing Context Acquisition under Cost, Privacy and ConsequenceAI Control SystemsModel-Risk Evidence for Agentic Systems: From Outcome Accuracy to Trajectory AssuranceDecision-Grade Agentic SystemsTest-Time Compute in the Enterprise: When Reasoning Tokens are Worth MoneyResearchThe Reasoning Stack, 2026ResearchThe ROI Denominator Problem: Hours Saved are not Benefits CapturedThe Value LedgerWhat AI Strategy Means When Models Become CommoditiesApplied AI Strategy, Transformation and ConsultingWhat Scaling Laws Don't Tell YouResearch

Machine Intelligence

First principles of representation, control and search
What Would Count as Intelligence?Intelligence, computation and representationProblems, Tasks and EnvironmentsIntelligence, computation and representationComputation as State TransformationIntelligence, computation and representationAlgorithms, Heuristics and Learned ProceduresIntelligence, computation and representationRepresentation Before IntelligenceIntelligence, computation and representationFeedback, Control and the Cybernetic LoopIntelligence, computation and representationLearning, Reasoning and Search are not SynonymsIntelligence, computation and representationFrom Programme to AgentIntelligence, computation and representationSearch as Structured ExplorationSearch, planning, logic and reinforcementHeuristics, a* and the Price of GuidanceSearch, planning, logic and reinforcementAdversarial Search and Strategic IntelligenceSearch, planning, logic and reinforcementPlanning Before ActionSearch, planning, logic and reinforcementKnowledge Representation and Logic: The Price of ExplicitnessSearch, planning, logic and reinforcementReinforcement Learning and Sequential ChoiceSearch, planning, logic and reinforcementThe Bitter Lesson and the Neuro-Symbolic ReplySearch, planning, logic and reinforcementVectors as Directions, not ListsMathematical language for learning systemsMatrices as MachinesMathematical language for learning systemsEigenstructure and Persistent DirectionsMathematical language for learning systemsOptimisation, Approximation and GuaranteesMathematical language for learning systems

Consciousness

Machine consciousness research and assuranceOpen series map →
A Comparative Atlas of Possible MindsExperience, theories and candidate subjectsA Consciousness Assurance Case: Governing Machines We Cannot ClassifyMachine Consciousness ResearchBefore Experience: The Minimal ConstraintsExperience, theories and candidate subjectsBinding and the Candidate-Subject BoundaryExperience, theories and candidate subjectsBuilding an Attention Schema Without Mistaking Report for ExperienceExperience, theories and candidate subjectsCounting and Combinatorial ExplosionMathematical language and representationDoes the Substrate Matter, or the Causal Organisation?Experience, theories and candidate subjectsEmergence and the Ruler that Creates ItExperience, theories and candidate subjectsFirst-Person, Interpersonal and Third-Person EvidenceInquiry and epistemic disciplineFive Words People Confuse: Intelligence, Consciousness, Sentience, Selfhood and AgencyInquiry and epistemic disciplineFrom Observation to Explanation: Correlation, Causation and ConstitutionInquiry and epistemic disciplineFunctions and Compositional Behaviour: Why the Model is the Path, not the PileMathematical language and representationHow to Read a Frontier ClaimInquiry and epistemic disciplineIntegrated Information and the Causal Structure of Artificial SystemsExperience, theories and candidate subjectsInteroception and Grounding Why Embodiment Matters for Models of MindConsciousnessLevels, Boundaries and the Grain of ExplanationInquiry and epistemic disciplineMachine Attention and Human AttentionConsciousnessMachine Attention and Human Attention a Careful DisanalogyConsciousnessNumber, Measurement and RepresentationMathematical language and representationResearch Without Theory VictoryExperience, theories and candidate subjectsRights, Forks and Interrupted MindsExperience, theories and candidate subjectsSets, Logic and ProofMathematical language and representationThe Hard Problem and the Price of Every ExitExperience, theories and candidate subjectsThought Experiments as Instruments, not OraclesInquiry and epistemic disciplineThree Ledgers: Function, Mechanism and ExperienceInquiry and epistemic disciplineWhat Counts as Evidence for Another Mind?Inquiry and epistemic disciplineWhat Predictive Processing and Kashmir Shaivism Agree AboutConsciousness