Muhammad H.

I build AI products that have to work with real users, not just in a demo: RAG chatbots that answer from your own documents and say so when they don't know, voice agents that pick up the phone, and agents that handle the admin after every sales call. For the past 3.5 years I've been a team lead on Stammer AI, a white-label platform where agencies build chat and voice agents and resell them to their own clients. It serves 30,000+ users. Before that I was a team lead at Arbisoft. A FEW THINGS I'VE BUILT On Stammer AI, every agent gets its own knowledge index and only answers when the match to the customer's content is strong enough. Otherwise it hands off, so end customers don't get made-up answers. Voice agents run on Retell AI with Twilio numbers, with interruption handling and short "got it" replies so calls feel natural. For a sales coaching platform (Gold Selling Standards), we score every Zoom and Teams call on 12 sections, with a separate weighted rubric for discovery, demo and renewal calls. Every score has to quote the transcript lines it's based on. We also fixed the Microsoft Teams recording bot that kept dropping calls. On SalesEdgeAI, no bot joins the meeting. We pick up the finished Zoom or Teams recording, analyze it, and write the opportunity, notes, tasks and a follow-up email draft into Salesforce, HubSpot or Zoho. For Carnegie Mellon researchers (DARE), we built a no-code agent builder: set up an agent in a form, then chain agents on a drag-and-drop canvas. Institutions that can't send data to the cloud self-host the vector database. On Notifier, a small, cheap model reviews every Reddit, X and LinkedIn mention and returns yes or no with a reason, so users only get real leads and the AI bill stays small. HOW I WORK I code with Claude Code and Codex every day, often several sessions at once on separate branches: one building a feature while another writes tests. Nothing ships without automated tests, an AI review and my own review. Before a prompt goes live, I run it against a set of real questions and score the answers for accuracy and grounding. I track the cost of every AI request and send simple tasks to cheaper models. Stack: Python, Django, FastAPI, React, Next.js, TypeScript, PostgreSQL, Redis, AWS AI and LLMs: OpenAI GPT, Claude, Gemini, LangChain, LangGraph, LiteLLM, MCP, prompt engineering RAG: Pinecone, Weaviate, pgvector Voice: Retell AI, Twilio, Whisper, Deepgram AI automation and integrations: n8n, Salesforce, HubSpot, Zoho, WhatsApp, Stripe If you have an AI feature that works in testing but not with real users, or an idea you want to scope, send me a message with what you're building. I usually start with a small paid sprint so you can see how I work before committing to a bigger build.