Muhammad H.

𝐈 𝐛𝐮𝐢𝐥𝐝 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐚𝐧𝐝 𝐑𝐀𝐆 𝐬𝐲𝐬𝐭𝐞𝐦𝐬 𝐭𝐡𝐚𝐭 𝐠𝐞𝐭 𝐦𝐞𝐚𝐬𝐮𝐫𝐞𝐝 𝐛𝐞𝐟𝐨𝐫𝐞 𝐭𝐡𝐞𝐲 𝐬𝐡𝐢𝐩, 𝐧𝐨𝐭 𝐚𝐟𝐭𝐞𝐫 𝐭𝐡𝐞𝐲 𝐛𝐫𝐞𝐚𝐤. Last RAG rebuild went from 18% to 87% Hit Rate@5 and now serves 1M queries a month LangGraph, Python, FastAPI, AWS. Your agent worked in the demo. Three weeks into production nobody can explain why it gave a wrong answer, what it cost to run last week, or whether it's getting worse. That's what happens when evals, observability, and cost controls get added after launch instead of before. 𝐖𝐎𝐑𝐊 𝐈'𝐕𝐄 𝐒𝐇𝐈𝐏𝐏𝐄𝐃 ▸ Maintenance triage agent (LangGraph, multi-tenant) 2-4 hours of manual review per request, down to under 30 minutes. Agent now resolves 60-85% end to end with zero human touch. Manager capacity doubled, no new hires. Shipped with a 14-grader eval harness and CI regression gates. ▸ Support RAG rebuild Hit Rate@5 18% → 87%. MRR@5 0.31 → 0.73. P95 time-to-first-token 3.2s → 1.9s with Redis semantic caching. Cost per correct answer dropped 15x. Root cause was a vocabulary gap between how customers ask and how the KB was indexed. Fixed at the source, not with a bigger model. ▸ Multi-tenant AI helpdesk One system, many client companies, zero data bleed. Per-tenant vector namespacing, token budgets, tool permissions scoped to each company's policy. Six adversarial attack tests run in CI and all must fail before code merges. Isolation enforced by infrastructure, not filtered by prompt. 𝐖𝐇𝐀𝐓 𝐈 𝐁𝐔𝐈𝐋𝐃 1. AI Agents & Multi-Agent Systems LangGraph agents, multi-step tool calling, MCP servers, human-in-the-loop approval gates, durable execution with Temporal so a crashed agent resumes instead of re-running and paying twice. 2. RAG & Retrieval Pipelines Hybrid retrieval, reranking, chunking strategy, vector DB selection, semantic caching, streaming. Retrieval that finds the right context consistently, and you have the numbers to prove it. 3. LLM Evaluation & Observability Eval harness built before the agent ships. Deterministic graders plus calibrated LLM-as-judge (took one from TPR 90%/TNR 53% to 95%/93%). Langfuse or LangSmith traces on every run. Deploys gated by evals, not by intuition. 4. Guardrails & Production Hardening Deterministic checks around every model call. PII redaction, RBAC, token budgets, cost caps, adversarial testing. Built for teams working under EU AI Act or similar. 5. Backend & Cloud FastAPI on AWS ECS, Lambda, Bedrock. Postgres with row-level security, Docker, CI/CD. Seven years of backend and cloud engineering before I moved to AI. 𝐒𝐓𝐀𝐂𝐊 Agents: Python · LangGraph · LangChain · OpenAI Agents SDK · Claude Agent SDK · Pydantic AI · AWS Strands · Temporal · MCP · A2A RAG: Qdrant · Pinecone · pgvector · Redis · Haystack · hybrid search · reranking · semantic caching Evals: RAGAS · DeepEval · Langfuse · LangSmith · LLM-as-Judge · pytest · GitHub Actions Infra: FastAPI · AWS ECS · Lambda · Bedrock · Docker · PostgreSQL · Terraform · LiteLLM Frontend: React · Next.js · TypeScript AWS Community Builder · AWS Agentic AI Certified · 7+ years cloud and backend before AI 𝐇𝐎𝐖 𝐈 𝐖𝐎𝐑𝐊 Most AI requests arrive already solved. "We need a RAG chatbot." "We need an agent for support tickets." I don't start there. I start one level up: what business outcome breaks if this doesn't exist, who feels it, and what does a correct result actually look like. Then I decompose backward into the system. That reordering is usually where the money is. The support RAG rebuild above was scoped as a model upgrade. The actual failure was a vocabulary gap between how customers phrase questions and how the KB was indexed. A bigger model would have cost more and fixed nothing. 𝐖𝐇𝐀𝐓 𝐇𝐀𝐏𝐏𝐄𝐍𝐒 𝐍𝐄𝐗𝐓 1. You describe the workflow that's stuck, not the solution you have in mind 2. Within 24 hours you get the problem decomposed: what actually breaks, what success would have to measure, and what one resolved task will cost at your volume 3. If it's worth building, we scope a first milestone with a defined eval gate Step 2 costs you nothing and is useful whether or not you hire me.