https://rajeshrmahapatra.com Articles and research writing on agentic AI, governance, evaluation, context and production architecture.https://rajeshrmahapatra.com/articles/deployment/real-time-fraud-investigation-agents.html https://rajeshrmahapatra.com/articles/deployment/real-time-fraud-investigation-agents.html A three-speed architecture for combining payment-time fraud controls, bounded agentic investigation and accountable customer-case decisions without placing a generative model directly in the authorisation path.https://rajeshrmahapatra.com/articles/deployment/kyc-periodic-review-evidence-bound-workflow.html https://rajeshrmahapatra.com/articles/deployment/kyc-periodic-review-evidence-bound-workflow.html A bank-grade design for customer due diligence that combines event-driven review, a scheduled control backstop, provenance-aware evidence assembly and explicit human authority over risk-rating and relationship decisions.https://rajeshrmahapatra.com/articles/deployment/aml-alert-triage-hard-sar-boundary.html https://rajeshrmahapatra.com/articles/deployment/aml-alert-triage-hard-sar-boundary.html A controlled architecture for using agents to assemble, test and present AML alert evidence while reserving suspicious-activity reporting, confidentiality and consequential relationship decisions to authorised people and policy.https://rajeshrmahapatra.com/articles/architecture/agent-identity-delegated-authority-three-principals.html https://rajeshrmahapatra.com/articles/architecture/agent-identity-delegated-authority-three-principals.html An enterprise agent is not its user and not its runtime service account. A control model for separating agent identity, human entitlement and tool authority before a consequential action is allowed to execute.https://rajeshrmahapatra.com/articles/architecture/context-architecture-reconstruction-over-accumulation.html https://rajeshrmahapatra.com/articles/architecture/context-architecture-reconstruction-over-accumulation.html A bank-grade context architecture that reconstructs current state, enforces authorization before retrieval, links claims to evidence and treats memory writes as governed actions rather than unlimited conversation history.https://rajeshrmahapatra.com/articles/architecture/the-certainty-gradient-placing-workflow-steps.html https://rajeshrmahapatra.com/articles/architecture/the-certainty-gradient-placing-workflow-steps.html A production architecture for decomposing banking workflows into deterministic controls, calibrated models and bounded reasoning, so autonomy is earned by evidence rather than granted by default.https://rajeshrmahapatra.com/articles/deployment/the-eval-pipeline-is-the-product.html https://rajeshrmahapatra.com/articles/deployment/the-eval-pipeline-is-the-product.html Prompts are disposable. Models are rented. The durable asset in an LLM system is the evaluation harness — the machinery that tells you, automatically and before your users do, whether the system still works.https://rajeshrmahapatra.com/articles/deployment/shipping-llm-systems-in-regulated-environments.html https://rajeshrmahapatra.com/articles/deployment/shipping-llm-systems-in-regulated-environments.html Getting 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 to production sign-off.https://rajeshrmahapatra.com/articles/aws/the-well-architected-agentic-stack.html https://rajeshrmahapatra.com/articles/aws/the-well-architected-agentic-stack.html A practitioner reading of AWS's dedicated Agentic AI Lens, revised 10 June 2026, applied to traces, identity, tool policy, memory, evaluation and unit economics.https://rajeshrmahapatra.com/articles/design/knowledge-architecture-as-design-discipline.html https://rajeshrmahapatra.com/articles/design/knowledge-architecture-as-design-discipline.html Your organization's knowledge now has two audiences: people and machines. Most companies have organized it for neither. Knowledge architecture is the design discipline that fixes this — and it is design, not plumbing.https://rajeshrmahapatra.com/articles/design/designing-interfaces-for-human-agent-teams.html https://rajeshrmahapatra.com/articles/design/designing-interfaces-for-human-agent-teams.html Agents don't fail because their models are weak. They fail because the interface between human and agent was never designed. Five patterns I use to build that interface deliberately.https://rajeshrmahapatra.com/articles/architecture/context-is-the-new-compute.html https://rajeshrmahapatra.com/articles/architecture/context-is-the-new-compute.html Model quality stopped being the bottleneck a while ago. What separates agentic systems that work from ones that flail is context engineering — and most teams still treat it as an afterthought to prompting.https://rajeshrmahapatra.com/articles/architecture/anatomy-of-a-production-multi-agent-system.html https://rajeshrmahapatra.com/articles/architecture/anatomy-of-a-production-multi-agent-system.html The 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 regulated bank — and why every one of them is load-bearing.https://rajeshrmahapatra.com/articles/aws/bedrock-agents-in-production.html https://rajeshrmahapatra.com/articles/aws/bedrock-agents-in-production.html A current field guide to AgentCore Harness and Runtime, Gateway, Identity, Memory, Cedar Policy, Evaluations and the controls required in a regulated production estate.https://rajeshrmahapatra.com/articles/machine-intelligence/what-would-count-as-intelligence.html https://rajeshrmahapatra.com/articles/machine-intelligence/what-would-count-as-intelligence.html Intelligence is a causal claim: a bounded system acquired transferable competence under controlled novelty, using declared priors, evidence and resources.https://rajeshrmahapatra.com/articles/research/what-scaling-laws-dont-tell-you.html https://rajeshrmahapatra.com/articles/research/what-scaling-laws-dont-tell-you.html Scaling laws predict loss, not capability. Four blind spots (data quality walls, post-training returns, architecture bets, evaluation ceilings) matter more to what ships in 2026 than the curve.https://rajeshrmahapatra.com/articles/consciousness/predictive-processing-kashmir-shaivism.html https://rajeshrmahapatra.com/articles/consciousness/predictive-processing-kashmir-shaivism.html Predictive processing and the Pratyabhijna school of Kashmir Shaivism both treat the ordinary given world as constructed rather than found, but they are built to answer different questions, and the popular claim that one proves the other is a specific, nameable error.https://rajeshrmahapatra.com/articles/consciousness/what-counts-as-evidence-for-another-mind.html https://rajeshrmahapatra.com/articles/consciousness/what-counts-as-evidence-for-another-mind.html Testimony is a causal trace whose evidential force depends on how the report was generated and which rival generators remain live.https://rajeshrmahapatra.com/articles/research/what-ai-strategy-means-when-models-become-commodities.html https://rajeshrmahapatra.com/articles/research/what-ai-strategy-means-when-models-become-commodities.html When capable models are broadly purchasable, durable advantage moves to the rights, routines and feedback loops that turn intelligence into better decisions and verified outcomes.https://rajeshrmahapatra.com/research/research/what-agent-benchmarks-actually-measure.html https://rajeshrmahapatra.com/research/research/what-agent-benchmarks-actually-measure.html An agent benchmark score is a property of a model, scaffold, tool environment, verifier, retry budget and task sample together: not a portable estimate of production utility.https://rajeshrmahapatra.com/research/research/web-computer-use-agents-grounding-recovery.html https://rajeshrmahapatra.com/research/research/web-computer-use-agents-grounding-recovery.html Once an agent can perceive a usable interface, reliable completion depends on binding actions to current state, detecting change and recovering without exceeding delegated authority.https://rajeshrmahapatra.com/articles/gcp/vpc-service-controls-workload-identity-agentic-estates.html https://rajeshrmahapatra.com/articles/gcp/vpc-service-controls-workload-identity-agentic-estates.html The security perimeter is not a bolt-on to an agentic platform. This architecture field guide places VPC Service Controls, workload identity federation, tool scope, encryption and egress around a bank-grade Google Cloud agent estate.https://rajeshrmahapatra.com/articles/machine-intelligence/vectors-as-directions-not-lists.html https://rajeshrmahapatra.com/articles/machine-intelligence/vectors-as-directions-not-lists.html A vector's coordinates change with the basis, while its useful geometry survives only under declared transformations, metrics and normalisation.https://rajeshrmahapatra.com/research/research/tool-use-as-learned-api-interaction.html https://rajeshrmahapatra.com/research/research/tool-use-as-learned-api-interaction.html Language models can learn when and how to call APIs, but a learned calling policy cannot establish authority, validate side effects or prove that the intended business state was reached.https://rajeshrmahapatra.com/articles/consciousness/three-ledgers-function-mechanism-experience.html https://rajeshrmahapatra.com/articles/consciousness/three-ledgers-function-mechanism-experience.html A claim about what a system does cannot become a claim about how it works or what it feels without an explicit bridge that can fail.https://rajeshrmahapatra.com/articles/consciousness/thought-experiments-as-instruments-not-oracles.html https://rajeshrmahapatra.com/articles/consciousness/thought-experiments-as-instruments-not-oracles.html A thought experiment earns force by exposing dependencies among premises, bridges and countermodels, not by making one intuition feel decisive.https://rajeshrmahapatra.com/articles/research/the-roi-denominator-problem-hours-saved-are-not-benefits-captured.html https://rajeshrmahapatra.com/articles/research/the-roi-denominator-problem-hours-saved-are-not-benefits-captured.html A practical method for converting AI time savings into schedulable capacity, verified operating outcomes and finance-owned value, without treating every minute returned as cash.https://rajeshrmahapatra.com/articles/research/the-reasoning-stack-2026.html https://rajeshrmahapatra.com/articles/research/the-reasoning-stack-2026.html Chain-of-thought, tool use, search, RL with verifiable rewards, test-time compute: five layers everyone name-drops and few can rank by evidence. Here's my map of what actually holds up and where the seams are.https://rajeshrmahapatra.com/articles/consciousness/the-hard-problem-for-engineers.html https://rajeshrmahapatra.com/articles/consciousness/the-hard-problem-for-engineers.html A decision map of the hard problem that credits what each exit explains, then records the assumptions, consequences and unresolved work it inherits.https://rajeshrmahapatra.com/articles/deployment/evaluation-stack-for-agentic-systems.html https://rajeshrmahapatra.com/articles/deployment/evaluation-stack-for-agentic-systems.html A practical guide to building golden case sets, adversarial scenarios, trajectory-evidence scoring and CI gates for regulated production environments.