Agentic AI infrastructure for production
I've been engineering agentic AI systems since January 2025: agents that take work from any channel, reason over your SOPs and history, act in your systems through secure tool layers, and stay inside the guardrails you set. I build them as infrastructure, not demos, with the same five-layer architecture every time so they're predictable, testable and easy to audit. I hold Anthropic's Claude Certified Architect and Developer certifications.
- 5 layersinterface, reasoning, memory, tools, guardrails
- 150+apps an agent can act in
- Model-agnosticClaude, GPT or open models, picked per task
- 4 weeksfrom survey to production
The agentic AI stack I build
Interface layer
Agents that take work wherever it arrives: chat, email, voice, webhooks or API, with one consistent core behind every channel.
Reasoning and orchestration
Multi-step and multi-agent workflows with planning, routing between specialised agents, queues, retries and human checkpoints, using the right model for each task.
Memory and retrieval
Retrieval over your SOPs, documents, CRM history and past decisions, so agents answer from your knowledge and stay consistent over time.
Tools and MCP servers
Secure, permissioned actions across HubSpot, Stripe, Slack, helpdesks, databases and 150+ apps, including custom MCP servers for internal systems.
Guardrails and governance
Approvals, spending and action limits, confidence thresholds and a full audit log, so a person signs off on anything risky.
Evals and observability
Test suites built from real cases, run before every change, plus tracing, outcome dashboards and cost monitoring in production.
Where it runs in production
- Inbox, ticket and form triage across support and operations
- Lead qualification, enrichment and CRM updates for sales teams
- Document intake: invoices, receipts, claims and forms extracted and filed
- Internal knowledge agents over SOPs, policies and past decisions
- Voice and chat agents sharing one reasoning, memory and tool layer
- AI features inside your own product, behind your auth and permissions
How it works
- Step 1
Week 1 · Survey
Map every workflow and call type, measure the hours it costs and set an ROI baseline.
- Step 2
Week 2 · Spec
Agent spec, data access, integrations and guardrails agreed in writing, with a fixed quote.
- Step 3
Week 3 · Build
Built in your stack and run against a test suite of real cases. You see it working this week.
- Step 4
Week 4 · Commission
Go-live, team training, monitoring dashboard and two weeks of close support.
Tools I use
- Claude API
- Claude Agent SDK
- MCP
- OpenAI
- Open models
- RAG
- Vector search
- Evals
- Tracing
- Node.js
- TypeScript
- Python
- n8n
Related work
Questions
What is agentic AI?
Agentic AI describes AI systems that act, not just answer: they decide which steps to take, call tools, check the results and continue until the job is done, with guardrails and human approval for risky actions.
What makes an AI agent production-ready?
A fixed architecture, clear permissions, an evaluation suite built from real cases, confidence thresholds with human hand-off, full logging and monitoring after launch. Without those, an agent is a demo.
What is MCP?
The Model Context Protocol (MCP) is an open standard for connecting AI models to tools and data. An MCP server exposes your systems to Claude and other AI clients in a controlled, permissioned way.
Which AI models do you use?
Whichever fits each task: Claude, GPT or open models. Reasoning-heavy steps, fast classification and private data each have different best choices, and the architecture lets the model change without rebuilding the agent.
What does a Claude Certified Architect do?
Anthropic's Claude Certified Architect certification covers designing production systems on Claude: agent architecture, tool use, evaluation and guardrails. Sumedh holds both the Architect and Developer – Foundations certifications.
How do you stop an AI agent from making mistakes?
By engineering for them: evals from real examples, limits on what the agent may do, confidence thresholds, full audit logs and a review queue so a person approves anything risky.
Sumedh
Open to freelance builds worldwide, full-time roles and nerdy chats about AI agents.