Sashi Kumar
I design production-grade agentic systems, AI platforms and cloud-native data architectures — from identity, context and memory to MCP, governance, observability and scale.
Enterprise systems I have designed — the patterns, boundaries and production considerations behind them.
AI-native data engineering with lineage-aware automation, governed transformation and human-controlled deployment.
Agentic Data Engineering · Knowledge Graphs · Lineage Intelligence · Metadata AI
Connecting requirements, code, tests and evidence through AI-driven software lifecycle intelligence.
Software Agents · Requirements Intelligence · Traceability · Test Generation
Identity-aware engineering knowledge retrieval across millions of controlled technical documents.
Enterprise RAG · Fine-Grained Authorization · Engineering Knowledge · Hybrid Search
Transforming complex enterprise documents into structured, searchable and decision-ready knowledge.
Intelligent Document Processing · Multimodal AI · Semantic Extraction · Confidence Routing
Identity-driven control of agent execution, tool access, budgets, concurrency and enterprise AI consumption.
Agent Governance · Multi-Tenancy · Policy Enforcement · MCP Security
Defense-in-depth controls across identity, context, tools, memory, data and autonomous execution.
Agent Security · Prompt Injection Defense · Tool Authorization · Memory Security
Deep dives into the parts of enterprise AI that only become interesting in production.
Why multi-tenant agent platforms should use one trusted identity for real-time rate limiting and cost attribution.
Enterprise systems and architecture problems I've worked on. Described by sector and scale — the customers are confidential, the architecture is the point.
Fortune 50 · Aerospace
Millions of engineering documents across decades of programmes, where who may read what is itself part of the domain.
Architecture and production readiness, discovery through delivery.
Fortune 10 · Technology
Onboarding many developers onto a data estate nobody could hold in their head, where the metadata was more valuable than the data.
Platform architecture and the agent/human boundary.
Digital Analytics Platform
Analytics for people who will not write SQL, against a schema where a wrong join is a wrong business answer rather than an error.
End-to-end architecture, evaluation strategy and guardrails.
Working explorations of enterprise AI architecture patterns — the implementations behind the writing.
How the writing connects — domains, and the areas within them.
Positions rather than preferences — each one is something a system either does or does not do.
Identity defines the execution boundary.
An agent's access to data, memory, and tools should derive from a verified user, tenant, and purpose.
Context is a governed dependency.
Useful context needs provenance, version, permissions, and freshness. What the system retrieves shapes what it can safely conclude.
Autonomy should be explicitly scoped.
Agents can plan and propose, while tool permissions, execution environments, budgets, and write actions remain enforceable.
Every decision should leave evidence.
Capture identity, source versions, retrieved context, tool calls, approvals, and resulting artifacts so outcomes can be explained.
Approval belongs at the point of consequence.
Require human review when an action grants access, merges code, publishes knowledge, or changes production state.
Deterministic controls surround probabilistic reasoning.
Use models for interpretation and synthesis; use explicit rules for authorization, validation, and deployment.
My focus is not getting an LLM demo working. It is the identity, context, memory, security, governance, observability and distributed-system concerns required to operate AI systems in production — and what it takes to move from prototype to something a business can depend on.