daimarh

👋 About me

I've spent 15+ years making commercial enterprise software trustworthy, first as a System Verification Tester at IBM, then as a Senior Software Engineer at HCLTech building test frameworks, and release quality for commercially shipped enterprise products.

Now I bring that same discipline to AI systems: agents that call real tools, follow business rules and are tested like production software, not demos.

🔭 Building agentic AI apps with LangGraph, Claude API and MCP
🧪 Testing LLM output the way I test APIs: assertions, edge cases, zero-escape goals
🌱 Learning: RAG patterns, LLM evaluation, cloud-native deployment
💬 Ask me about: test strategy, CI/CD speed-ups, root-causing hard defects

🚀 Featured projects

🤖 Support Agent: LangGraph + MCP

Multi-agent ticket resolution

A 5-node LangGraph StateGraph (triage → context → resolver → escalation → finalizer) that resolves support tickets end to end. Business rules live in the graph's topology, not just the prompt.

✅ 60% auto-resolved at a 0.7 confidence threshold
🛡️ Capped at 8 node-visits to guarantee termination
🧪 16 passing tests

🎫 support-agent-mcp

AI "Resolve with AI" for support desks

Agents click Resolve with AI and Claude calls 4 MCP tools (get_ticket, lookup_customer, check_inventory, update_ticket_status). It returns a drafted customer reply plus a step-by-step reasoning trace in one round trip.

🌐 Next.js 14 on Vercel → FastAPI on Railway
🔐 Supabase Auth (JWT) + Postgres with RLS
🧰 TypeScript MCP server over stdio JSON-RPC

💬 Natural-Language-to-SQL

Ask your database in plain English

A full-stack app that turns plain English into executable SQL via the Claude API, with generated SQL checked for correctness and result accuracy.

🧩 4 integrated services: Next.js, FastAPI, Supabase Postgres, Supabase Auth
🧪 Playwright end-to-end + pytest API suites
🎯 Zero defect escapes across 12 build iterations

🔁 cicd-flask-lab

Spec-driven CI/CD, end to end

Every pipeline stage is owned by a declarative spec file. The code must conform to the spec, not the other way around.

🧱 lint → test (≥80% coverage gate) → SAST → build → Trivy CVE gate → deploy → verify
🔒 Security: bandit, pip-audit, gitleaks
☸️ Terraform, Helm, Minikube, Prometheus

🧰 More: MCP Document Tools server

A Python package that exposes document conversion and processing tools through an MCP server, so AI assistants can call them directly. Tools are typed Python functions with Pydantic Field descriptions for clean tool schemas.

🛠️ Skills dashboard

🤖 AI / LLM
💻 Languages
🧪 Testing
🔁 CI/CD & DevOps
🌐 APIs & Web
🗄️ Data
🤝 Practice

💡 How I work

"If it isn't tested, it isn't done, and that includes the AI."

Principle	In practice
🎯 Rules in structure, not hope	Agent guardrails enforced by graph topology and hard caps, not only by prompts
🧪 Test the output, not just the code	Assert generated SQL and LLM answers for correctness, like API contracts
⚡ Fast feedback wins	Parallel pipelines and auto-ticketing so defects surface in minutes, not days
🔍 Find the real root cause	Binary-search isolation + log correlation before any fix ships

Let's build AI that people can trust.
LinkedIn · Email

Navigation Menu