Guelph, Ontario · available for advisory work, speaking, and collaboration

Anirban Kanjilal

Solutions Architect · Cloud · Security · AI · Quantum

I build the AI platforms, cloud architecture, and security that run in production, with a research track in quantum.

I am a systems architect first. My foundation is more than twenty years of designing and engineering high-volume, resilient platforms for regulated industries such as payments, where downtime and data loss are unacceptable. That depth in system architecture and platform engineering is what I bring to every engagement, whether the workload is a payments switch, a cloud migration, or an AI platform.

Today, much of my work is helping organizations take artificial intelligence from proof of concept to regulated, always-on production. That work rests on the same fundamentals: sound cloud architecture, disciplined platform engineering, and security built in from the start. The model is rarely the difficult part. The difficulty is meeting the security, cost, resilience, and service level commitments that production demands.

My quantum work covers quantum machine learning, quantum algorithms, materials for quantum sensors, and post-quantum cryptography. The last of these is the most immediate, because the public-key algorithms protecting payments today must be migrated to the standards NIST finalized in 2024 before a large quantum computer can break them.

View the work My blogs GitHub

Focus

Machine learning in production

Quantum on Amazon Braket

Artificial intelligence security

Cloud and AI cost engineering

Agentic systems

Selected work

enterprise-ai-archlive demo

A complete nine-layer enterprise AI architecture that deploys into a real AWS account with a single action. One governed request passes through all twelve hops, reaching the model at step seven, after which the environment is torn down automatically. It uses retrieval on S3 Vectors, Bedrock Guardrails that stop a poisoned document at the retrieval boundary, a Cedar validation checkpoint, and an immutable audit record for every run.

awsbedrockgovernance

Try the demo · Read the article

gha-ecs-runnerlive demo

Self-hosted GitHub Actions runners on ECS Fargate that build container images privately and push them to Amazon ECR. The registry is locked so that only the runner can push, and every image is signed within the account. Each task runs a single job, produces one artifact, and is then destroyed. The repository includes the full Terraform, the runner image, and the control plane.

awsci-cdsupply-chain

Try the demo · Read the code

llm-security-lablive demo

A working demonstration of prompt injection rather than a written description of it. It provides vulnerable and hardened versions of common language model risks, each mapped to the OWASP Top 10 for Large Language Model Applications, MITRE ATLAS, and the NIST framework. It runs entirely offline and requires no API keys.

securityowaspatlas

Try the demo · Read the code

quantum-on-braketlive

Quantum algorithms that run on Amazon Braket and are written with cost in mind. They run for free on the local simulator and can be pointed at real hardware with a single flag, and a spending guard runs before every call to a quantum processor. Grover's algorithm is available now, and an optimization algorithm is next.

quantumbraketoptimization

Run it

ai-finops-agentlive demo

An agent that reads billing data in the open FOCUS format and shows where cloud and model spending goes, from idle accelerators to sudden increases in usage. The agent itself is hardened against prompt injection.

costbillingagents

Try the demo · Read the code

llm-cost-guardrailslive demo

A circuit breaker for language model spending. This provider-neutral middleware meters every call and rejects any request that would exceed budget, using per-tenant budgets, rate limits, and a spending kill switch. It runs offline and requires no API keys.

costfinopsguardrails

Try the demo · Read the code

mcp-secure-gatewaylive demo

A hardened gateway that sits in front of Model Context Protocol tools. It authenticates the caller, enforces a default-deny tool allow list, validates input, and treats tool output as untrusted data, so that a confused deputy attack is blocked at the boundary. The audit log records every decision.

mcpagentszero-trust

Try the demo · Read the code

agent-eval-harnesslive demo

Unit tests for agent behavior. This framework-neutral harness verifies that an agent calls the correct tool, refuses an attack, does not leak secrets, and stays within a token and cost ceiling. It then gates deployment on those results, in the same way a test suite gates a release.

agentsevaluationci

Try the demo · Read the code

On the roadmap

agentic-patterns reference-architectures

Reference architectures

EKS payments platformpayments

A multi-region, active-active Amazon EKS design for a critical payments platform. It uses a cell-based synchronous ledger to prevent data loss across two Canadian regions, aligns with PCI requirements, and includes the full detailed design in Terraform, Kubernetes, and CI/CD.

eksmulti-regionpayments

Read the architecture

Secure landing zonefoundation

The governed foundation that every other workload is deployed into. It is a multi-account, two-region organization built on AWS Cloud WAN, with service insertion, layered guardrails, and centralized inspection.

awsnetworkingguardrails

Read the architecture

Zero trust AI gatewayAI security

A single control point between agents and the models and tools they call. It provides scoped identity, an allow-listed tool registry, request and response validation, and a complete audit trail.

zero-trustagentsmcp

Read the architecture

Managed GHA runners on ECSdelivery

Self-hosted GitHub Actions runners implemented as ephemeral Fargate tasks. They build images without a daemon, sign them in the account with KMS, and push to ECR. The registry is locked so that the runner is the only writer, and each task runs a single job, produces one artifact, and is then destroyed.

ci-cdsupply-chainfargate

Architecture · Walkthrough

Secure RAG platformRAG

Private data grounding that never leaves the network. It provides managed retrieval and reranking, guardrails for prompt attacks and personal data, and continuous evaluation.

ragretrievalsecurity

Read the architecture

Resilient transaction corepayments

An idempotent, event-driven backbone for high-value transactions. It is designed against all six Well-Architected pillars and uses the saga pattern together with multi-region failover.

paymentsresiliencesaga

Read the architecture

Model lifecycle governanceMLOps

Training, registration, approval, deployment, and monitoring in one governed pipeline on SageMaker. It includes bias and explainability checks, drift detection, and the evidence trail that auditors require.

mlopssagemakergovernance

Read the architecture

Build and deploy pipelinedelivery

A single pipeline on GitHub Actions that takes a commit to production. It performs static and dynamic analysis, dependency and image scanning, signing and provenance, and keyless promotion across accounts. Releases are guarded, using GitOps for Kubernetes and blue-green deployment for the container service.

ci-cdsupply-chaingitops

Read the architecture

Background