I build autonomous AI agent systems that reason, remember, and act — from multi-agent deliberation and shared memory banks, to RAG pipelines, MCP tool servers, and LangGraph orchestration. My work sits at the intersection of LLM orchestration, AWS infrastructure engineering, and applied system design.

Currently: AWS Cloud Engineer at KodeKloud — designing and validating AI sentinel systems for multi-organisation AWS lab account governance.

AI Sentinel Ecosystem 5-agent autonomous system governing AWS lab accounts across multiple AWS Organisations — quorum deliberation, cost enforcement, predictive autoscaling, account recovery, and learner intelligence. Built on the Claude Agent SDK with a deterministic 20+ tool MCP layer. Validated across 49 runs, 30 scenarios: 98.4% detection accuracy, zero false-positive quarantines.

Claude Sandboxed Agent Production support intelligence system where Claude operates inside a strict tool sandbox — three explicitly permitted tools (account state, policy evaluation, audit trail), each with isolated credentials. Live infra access, not static RAG. WebSocket streaming, Firebase auth, and team-scoped JWT validation enforced at the path level.

AWS SAR MCP Server FastMCP server exposing the AWS Service Authorization Reference as LLM-callable tools — look up IAM actions, find destructive operations by service, compare permissions, and search across all 300+ AWS services. Zero AWS credentials required.

mem0 Pipeline Three-layer developer memory pipeline — mem0 fact extraction, Qdrant vector store, and Neo4j entity graph — with Claude reasoning over both layers for deep developer insight generation from GitHub activity signals. Contributed as a cookbook to mem0ai/mem0.

Dual Agent Memory Two Claude agents (AWS Resource Finder + AWS Recommender) sharing a persistent Hindsight memory bank via the Claude Agent SDK. Demonstrates tool use, agentic loops, and cross-agent memory patterns.

LangGraph Agent Stateful AWS infrastructure audit agent built with LangGraph and Claude. Conditional routing skips deep investigation on clean accounts (~40% cost reduction). Typed AuditState, MemorySaver checkpointing, 43 tests passing.

Bedrock RAG Pipeline End-to-end RAG pipeline on AWS Bedrock — document ingestion to S3, Knowledge Base creation, ingestion job orchestration, and grounded generation via the RetrieveAndGenerate API with Claude. Structured citations on every response.

AI & Agents

Cloud & Infrastructure

Backend & DevSecOps

• AWS Solutions Architect – Associate
• GCP Professional Cloud Architect
• HashiCorp Certified Terraform Associate
• TOGAF 10 – Level 1  |  Member, Association of Enterprise Architects

• Email: tanishka.marrott@gmail.com
• LinkedIn: linkedin.com/in/tanishka-marrott
• Blog: cloud-design-diaries.hashnode.dev