Muhammad T.

Most AI projects fail for the same reason: the agent sounds smart but doesn't actually know anything about your business. I build the other kind which is, agents grounded in your real data, your documents, and your tools. I'm a software engineer with 8+ years across full stack development, DevOps, and for the last stretch, production AI agent systems. My work sits at the point where those three overlap, which turns out to be exactly where most AI projects get stuck: the model is the easy part, and the data pipelines, permissions, infrastructure, and safe execution are the hard part. - What I've been building lately ▸ LumenAI: an AI platform for private equity, VC, and family office firms. I built a suite of specialist agents that turn raw company documents into diligence-ready output: company memos and primers, financial and operational metric extraction, field extraction from unstructured PDFs, industry and competitive analysis, red flag and opportunity detection, P&L analysis, valuation support, and a grid execution agent that runs a set of KPI questions across dozens of companies or fiscal years and compiles the answers into a structured table. Analysts stopped reading 200-page data rooms manually. ▸ Ziggy (PPCEntourage): An insight and recommendation agent for Amazon sellers and PPC managers. Not a bolt-on chatbot: Ziggy reads page-level context from inside the application and uses an MCP server to pull additional account and campaign data on demand, then explains what's actually wrong with a campaign and what to do about it. ▸ Adyard Personal Assistant: An action-taking agent for Bol advertising. This one goes past recommendations: through a purpose-built MCP server, it retrieves account data, answers campaign questions, and performs user-requested actions inside the platform. An operating assistant, not a support bot. ▸ Enterprise AI Operations Agents: Internal workflow agents built around a company's actual processes rather than one generic assistant. Multiple specialist agents and sub-agents, each owning a workflow, so the system is testable and safe to expand. - How I approach it Start with insight and decision support. Move to automation only once the data, permissions, safeguards, and test coverage are strong enough that a mistake can't cost the client money. On the advertising projects specifically, that discipline is the whole ballgame, an agent that autonomously edits ad campaigns before it's been properly tested is a financial liability, not a feature. Phased rollout, human-in-the-loop first. - The engineering underneath Eight years of shipping production systems is what makes the AI work hold up. Before agents, I spent years on the unglamorous side: Jenkins and GitHub Actions pipelines, Dockerised and Kubernetes-orchestrated services on AWS, Apache Airflow for data orchestration, multi-tenant deployment automation, and full stack builds in Django, Laravel, React, Next.js, and Vue. I've mentored junior developers, run Agile teams, and handled direct client communication on UK and EU projects since 2019. - Core areas AI agents & LLM systems: Multi-agent orchestration, sub-agents, RAG, tool use, MCP servers Document intelligence: Summarization, field & metric extraction, structured output from unstructured data Backend & full stack: Python, Django, PHP, Laravel, Node.js, React, Next.js, Vue.js DevOps & cloud: Docker, Kubernetes, AWS (ECS/ECR), Jenkins, GitHub Actions, Azure DevOps, CircleCI Data & orchestration : Apache Airflow, PostgreSQL, MySQL, Redshift Business systems : ERPNext customization, subscription & financial modules, multi-tenant architecture - The trajectory 2018: Laravel and Vue modules for e-commerce clients. 2019–2022: owning CI/CD and mobile build automation, leading Agile teams, talking directly to clients. 2022–present: senior full stack and DevOps lead, then architecting AI agent platforms for investment firms and advertising SaaS products. Each layer built on the last, which is why I can take an AI project from "interesting demo" to "runs reliably in production," rather than handing off at the hard part. - Working together The clients who get the most out of me tend to arrive with a real business problem rather than a spec for a chatbot. "our analysts spend three days per deal reading documents," "our users can't tell why a campaign is underperforming," "this process eats twenty hours a week." If that sounds like your situation, send me the messy version of it. I'll tell you honestly whether an agent is the right answer, what the phased path looks like, and where the risks are before you spend anything. I'd rather scope it properly than sell you a demo.