Production AI Platforms & Agents on AWS
Move AI systems from prototype to reliable production.
InfraTales helps engineering teams make production decisions about agent permissions, evaluation, RAG, failure recovery, observability and cost on AWS. Led by Rahul Ladumor, Principal Cloud & AI Platform Architect.
Rahul works directly with your engineering team. Recommendations are prioritised by business impact, reliability, security, cost, maintainability and implementation effort.
Production AI architecture
When should an AI agent be allowed to act?
A useful agent needs access to tools. A production agent needs explicit boundaries around what it can read, propose and execute.
Decision context
- Workflow
- A customer-support agent can read an account, draft a response and propose a refund.
- Risk
- A refund changes a financial record. The same tool can also expose data across tenants if identity and scope are wrong.
- Decision
- Reads and drafts can proceed automatically. Refunds pass a deterministic policy check and require approval above the agreed threshold.
- Boundary
- The model proposes an action. A separate executor validates and performs it with a short-lived, least-privilege identity.
This is a reference scenario. The policy depends on the action's reversibility, financial impact, data scope and regulatory context.
Control path
- 01User requestIdentity and tenant established
- 02Agent proposesTool, arguments and reason
- 03Policy evaluatesScope, risk and approval rule
- 04Human approvesOnly when the action requires it
- 05Executor actsScoped credential and idempotency key
- 06Audit recordsProposal, decision and outcome
Architecture decisions
- The agent proposes; the executor acts.Model output never becomes a privileged API call without typed validation and an explicit policy decision.
- Approval is based on action risk.A low-risk read can be automatic. Irreversible, financial or cross-tenant actions need stronger controls or a human decision.
- Every attempt is observable and safe to retry.Record the proposal, policy result, approver and tool outcome. Use idempotency keys so a retry cannot repeat a side effect.
Read the analysis: when an AI agent should require human approval →
Engagements
- 01
Production AI Architecture Review
Your AI prototype works in testing. What evaluation, guardrails, cost and observability look like at real scale is still unclear.
How the review works → - 02
AWS Architecture Review
Your AWS architecture is growing harder to change, more expensive, or less reliable. A significant decision is in front of you.
How the review works → - 03
Fractional Principal Cloud & AI Architect
Your team makes architecture decisions regularly without a senior architect. Velocity is dropping or consistency is suffering.
How the engagement works →
Selected work
Rahul's cost and reliability audits have removed more than ₹50 lakh (about $60,000) a year in AWS waste for clients.
- 01AWS cost audits: up to 55% lower spendTwo-week AWS cost audits for fintech, e-commerce and SaaS clients: 47 idle EC2 instances removed, Savings Plans coverage 0% to 78%, spend down up to 55%.
- 02Enterprise RAG on Bedrock across 2M+ documentsA production RAG platform on Amazon Bedrock: answers in under 10 seconds instead of 30+ minutes, MRR@10 from 0.31 to 0.78, about 60% lower repeated LLM spend.
- 03KFC Thailand: PHP monolith to event-driven orderingHow KFC Thailand's ordering platform moved from a failing PHP monolith to event-driven services: P99 2.8s to under 200 ms, about 30% lower infra cost.
Insights
Is an AI system moving from demo to production?
Start with a short qualification form. Rahul reads every request and replies with a clear next step.