World IV
The Build
Deployment patterns, system design, AWS and Google Cloud, and the craft of working inside large organisations.
Cloud & Engineering
System design, AWS, Google Cloud and production practice. Open the shelf →
Written by Rajesh Ranjan Mahapatra
AWS & Agentic AIEngineering intelligent systems across clouds, tools and organizations
GCP and Agentic AIFrom accountable action to production ownership
Inference Engineering and Agentic AIFrom token physics to governed autonomous systems
Multi-Cloud Systems from First PrinciplesCapacity, state, failure and service boundaries
Quantization and Fast InferenceA visual first-principles book
Work & Careers
Judgement, influence and money, written for working people. Open the shelf →
Written by Rajesh Ranjan Mahapatra
Essays
61 essays, newest first where the site knows the date.
- Three Principals, One Decision: Agent Identity and Delegated Authority in Regulated SystemsArchitectureAn enterprise agent is not its user and not its runtime service account. A control model for separating agent identity, human entitlement an
- Agent Memory as an Attack Surface: Provenance, Quarantine, DriftArchitectureA single poisoned attachment wrote a false fact into an agent's long-term case memory, and the fix that followed treats every memory write a
- An Agentic Test Migration Factory with Repository GatesArchitectureA factory architecture for converting legacy test assets into maintainable code by separating discovery, intent recovery, generation, verifi
- Anatomy of a Production Multi-Agent SystemArchitectureThe demo takes a weekend. The platform takes a year. Here are the six layers that actually exist inside a multi-agent system running in a re
- Governed Context Architecture: Reconstruction, Provenance and Active Memory for Long-Horizon AgentsArchitectureA bank-grade context architecture that reconstructs current state, enforces authorization before retrieval, links claims to evidence and tre
- Context Caches under Banking Privacy ConstraintsArchitectureA threat-led architecture for caching prompts, retrieval results, context packets and tool reads without crossing customer, user, case, purp
- Context Compilers for Policy-Filtered RetrievalArchitectureA reference architecture for compiling identity, purpose, policy, time and evidence requirements into a bounded retrieval plan before an age
- Context is the New ComputeArchitectureModel quality stopped being the bottleneck a while ago. What separates agentic systems that work from ones that flail is context engineering
- Decision Decomposition in Commercial Credit: Evidence, Arithmetic, Policy, Narrative and AuthorityArchitectureA production architecture for separating document extraction, verified calculation, policy interpretation, narrative drafting and accountabl
- Designing Multi-Agent Topologies with Computable Blast RadiusArchitectureA method for treating agent tool permissions as a reachability graph so that, before deployment, you can compute the maximum damage any sing
- Event-Sourced Relationship Intelligence Without a Shadow Customer RecordArchitectureA federated architecture for assembling timely, cited relationship briefs across CRM, products, risk, pipeline, communications and public so
- Freshness, Supersession and Valid-Time EvidenceArchitectureA temporal evidence architecture for agents that must distinguish current, corrected, superseded and decision-time-valid facts instead of tr
- GraphRAG, Relational State or Vector Search?ArchitectureA workload-led method for separating documentary retrieval, governed case state and relationship traversal before selecting vector, relation
- Legal Document Investigation as Reconciliation, not SummarisationArchitectureA case-state architecture for extracting, comparing and reconciling parties, powers, restrictions, dates and inconsistencies across legal do
- Multimodal Context for Lending and Financial CrimeArchitectureA lineage-preserving architecture for combining scans, handwriting, tables, photographs and system records without turning extraction into f
- Deterministic-First Agents: The Certainty Gradient for Governed Banking AutomationArchitectureA production architecture for decomposing banking workflows into deterministic controls, calibrated models and bounded reasoning, so autonom
- Agentic Architectures on AWS: AgentCore, Durable Workflows, and Where Lambda EndsAWSA current map of AgentCore, durable workflow services and AWS compute primitives, including the boundary at which Lambda stops being the rig
- Amazon Bedrock AgentCore in Production: A Regulated-Industry Field ReportAWSA current field guide to AgentCore Harness and Runtime, Gateway, Identity, Memory, Cedar Policy, Evaluations and the controls required in a
- Cost Engineering Agentic Workloads on AWS: Bedrock Inference and AgentCoreAWSA measurement-led playbook for Bedrock inference, AgentCore Runtime, Gateway, Policy, Memory, Observability and Evaluations, without assumin
- Identity and Policy for Agents on AWS: AgentCore, Cedar, IAM, and Scoped AuthorityAWSA practical account of moving agent authorisation from one broad role to AgentCore workload identity, per-tool Cedar decisions, scoped AWS c
- Reading the AWS Well-Architected Agentic AI Lens in PracticeAWSA practitioner reading of AWS's dedicated Agentic AI Lens, revised 10 June 2026, applied to traces, identity, tool policy, memory, evaluatio
- Agent Reliability Atlas: 171 Production Failures, Controls and Recovery PathsDeploymentA field manual for diagnosing agentic AI, MCP, A2A, RAG, orchestration, cloud-runtime and model-serving failures by locating the first unsaf
- AML Alert Triage with a Hard SAR Boundary: Evidence Assembly Without Delegating SuspicionDeploymentA controlled architecture for using agents to assemble, test and present AML alert evidence while reserving suspicious-activity reporting, c
- Calibrated Cascade Routing in Production: Small Models First, Confident EscalationDeploymentA practical account of how tiered model cascades, calibrated confidence scoring and semantic caching can cut inference cost and latency whil
- Continuous Control Evidence: From Cloud Configuration to Auditor WorkpaperDeploymentA production design for turning approved cloud-control intent into read-only tests, reproducible evidence, exception queues and signed workp
- Drift: Behavioural Sentinels and Why Tool Selection Can Lead Output MetricsDeploymentA field guide to detecting behavioural drift in production agents by comparing observable tool, evidence and policy events with versioned ba
- The Evaluation Stack for Agentic Systems: Golden Cases, Adversarial Scenarios, and Process EvidenceDeploymentA practical guide to building golden case sets, adversarial scenarios, trajectory-evidence scoring and CI gates for regulated production env
- KYC Periodic Review as an Evidence-Bound Workflow: Events, Provenance and Human AuthorityDeploymentA bank-grade design for customer due diligence that combines event-driven review, a scheduled control backstop, provenance-aware evidence as
- Measuring Hybrid Retrieval Beyond Recall@kDeploymentAn evaluation system for lexical, dense, graph and reranked retrieval that measures authority, evidence coverage, contradictions, duplicates
- Population-Scale Outcome Testing for Customer InteractionsDeploymentAn evidence architecture for moving from small manual samples to risk-based testing across calls, chats and complaints while keeping custome
- RCSA Evidence Assembly Without Delegating the RatingDeploymentA controlled architecture for combining risks, controls, indicators, incidents, issues and prior assessments into a reviewable RCSA evidence
- Real-Time Fraud Investigation Agents: Fast Evidence, Bounded Action and Human AuthorityDeploymentA three-speed architecture for combining payment-time fraud controls, bounded agentic investigation and accountable customer-case decisions
- Regulatory Change as a Traceable Obligation GraphDeploymentA source-linked architecture for comparing regulatory versions, extracting candidate obligations, mapping them to policies and controls, and
- Shipping a Private Skill Model Foundry: Distillation, LoRA, Calibration, and Red Teaming Inside a VPCDeploymentA working account of how a bank builds, calibrates, and red-teams its own narrow skill models inside a closed network instead of calling a p
- Shipping LLM Systems in Regulated EnvironmentsDeploymentGetting an LLM system live in a bank is as much an evidence problem as a technology problem. This is a practical path from working prototype
- The Eval Pipeline is the ProductDeploymentPrompts are disposable. Models are rented. The durable asset in an LLM system is the evaluation harness: the machinery that tells you, autom
- Designing Human-Agent Handoffs for Regulatory ScrutinyDesignA control-focused examination of the moment an agent transfers evidence, decision rights and accountability to a qualified human reviewer.
- Designing Interfaces for Human-Agent TeamsDesignAgents don't fail because their models are weak. They fail because the interface between human and agent was never designed. Five patterns I
- Evidence-First UX: Showing Claims with Their ProvenanceDesignA wrong balance figure in a customer letter shows why generated claims need visible, per-claim provenance instead of a single buried sources
- Knowledge Architecture as a Design DisciplineDesignYour organization's knowledge now has two audiences: people and machines. Most companies have organized it for neither. Knowledge architectu
- The Autonomy Ledger: Making Agent Permissions LegibleDesignA governed, plain-language artefact that translates agent permissions into terms a risk committee can actually sign off on, so approval refl
- Building Bank-Grade Agents on Agent Runtime: LangGraph and a Controlled Path to ADKGoogle CloudAn architecture field guide for taking agentic workloads from a LangGraph prototype to Agent Runtime on Gemini Enterprise Agent Platform, wi
- BigQuery as the Agent's Structured Memory: Medallion Pipelines, CDC and Row-Level SecurityGoogle CloudVector indexes give an agent recall of unstructured text. BigQuery, built as a disciplined medallion pipeline with change data capture and r
- BigQuery as the Memory Layer for Agent PlatformsGoogle CloudMost agent memory debates obsess over vector databases. In a regulated enterprise, the analytical memory of your agent platform is BigQuery:
- Retrieval on GCP: Hybrid Dense and Sparse Search with RRF and Cross-Encoder RerankGoogle CloudWhy single-signal retrieval fails in enterprise settings, and how combining Vector Search on ScaNN with AlloyDB's sparse and structured sign
- Gemini Enterprise Agent Platform: What Actually Ships in a BankGoogle CloudGemini, Agent Runtime, ADK, grounding and VPC Service Controls form a credible regulated stack only when identity, location and feature-leve
- VPC Service Controls and Workload Identity for Agentic EstatesGoogle CloudThe security perimeter is not a bolt-on to an agentic platform. This architecture field guide places VPC Service Controls, workload identity
- Brownfield Intelligence: Coexisting With the Existing Enterprise | Rajesh MahapatraResearch essaysA ten-control coexistence envelope for coupling AI to brownfield enterprise systems without creating a second truth, divergent writes or per
- Career Architecture for Human-Agent Work | Rajesh MahapatraResearch essays
- Context Architecture Before Context Windows | Rajesh MahapatraResearch essaysA nine-stage context supply chain, explicit context budget and interactive router for assembling authorised enterprise evidence at runtime.
- Decision Rights as the Hidden Architecture of AI Strategy | Rajesh MahapatraResearch essaysA practical strategy paper on compiling decision rights into AI operating systems, with a live Decision-Rights Compiler.
- Interoperability Across MCP, A2A and Enterprise Events | Rajesh MahapatraResearch essaysAn interoperability envelope for choosing API, MCP, A2A, enterprise event or human workflow boundaries without losing semantics, authority o
- Leadership Capability in the Agentic Enterprise | Rajesh MahapatraResearch essaysA chief executive tells every function to reach sixty per cent AI adoption before year end. Licences spread, teams demonstrate agents and da
- Making Unstructured Enterprise Knowledge Operational | Rajesh MahapatraResearch essaysAn eleven-field Operational Knowledge Assertion and interactive router for transforming governed documents into evidence, assertions and wor
- Test-Time Compute in the Enterprise: When Reasoning Tokens are Worth MoneyResearch essaysSpending more compute at inference time, rather than only at training time, is now a deliberate lever an enterprise team can pull. It is als
- The New Role System Around Enterprise AI | Rajesh MahapatraResearch essaysAn event-level role system for allocating AI work, authority, evidence, interfaces, capacity, challenge, transition and role expiry without
- The Shadow Agent Network Inside the Enterprise | Rajesh MahapatraResearch essaysDiscover and contain hidden agent networks by tracing identity, data, tools, authority, business effects and persistent state rather than mo
- What Scaling Laws Don't Tell YouResearch essaysScaling laws predict loss, not capability. Four blind spots: data quality walls, post-training returns, architecture bets, evaluation ceilin
- When an Enterprise Knowledge Graph Earns Its Cost | Rajesh MahapatraResearch essaysA four-book Graph Dividend Ledger and interactive investment router for testing whether governed enterprise relationships justify their life
- When Change Management Becomes Architecture | Rajesh MahapatraResearch essaysA change-bearing architecture that embeds proficiency, discretion, support, feedback, workload and workforce transition in consequential hum
- Who Should Own the Enterprise AI Control Plane | Rajesh MahapatraResearch essaysA decision-rights charter for allocating enterprise AI control-plane authority across platform, domain, specialist, executive and independen
Research papers
7 papers from the research programmes.
- Brain architecture as a changing networkWhy the brain's functional architecture changes without rewiring: a first-principles account of hierarchy, recurrence, oscillatory timing, s
- Hybrid temporal architectures and the thick presentA first-principles architecture and executable benchmark for deciding when attention, recurrence and state-space memory genuinely preserve d
- Motivation, goals and veto architectureWhy action depends on a timed contest between drives, habits, goals and vetoes, and how to redesign that contest before willpower is needed.
- One architecture, many maps?A first-principles method for comparing Indian, Buddhist, Chinese and phenomenological maps of mind without forcing equivalence, with an exe
- Proof of governed execution and the five-chain ledgerA practical first-principles design for proving that a consequential digital action was performed by the right identity, on the right eviden
- Scaling laws, test-time compute and apparent emergenceA first-principles account of why capability cliffs can be created by metrics, multi-step tasks and test-time search, with an executable aud
- Three ledgers: function, mechanism and experienceA practical method for stopping claims about what an AI system does from being mistaken for claims about how it works or what it may experie
Insight Editions
4 study editions with covers, PDFs and EPUBs. All editions →
