Most AI demos work. Far fewer survive their first month with real users. I have spent eight years on the gap between the two.
Google, Adobe, Wayfair, Virgin Media, most recently Solventum (3M Healthcare). Today that means agent graphs in LangGraph, retrieval that can cite its sources, and the evals and cost controls that decide whether a feature ships — on top of eight years of classical ML and data engineering.
I write about it on Medium. If you are building something in this space, I would like to hear about it.
How I think about an AI system, layer by layer — and the tools I use at each one.
Build with it constantly Shipped to production Working knowledge
Orchestration, tool use and the human checkpoints
Getting the right context in, and proving it was the right context
Language models, and the classical models underneath them
Traces, evals and cost ceilings, so nothing regresses quietly
Where all of it actually runs
The foundation everything above depends on
8+ years building production AI systems — from classical ML and deep learning to Agentic AI and RAG pipelines. Track record of shipping scalable models with full MLOps (CI/CD, Docker, Kubernetes, MLflow) on AWS and GCP for companies processing millions of daily transactions.
Healthcare technology company spun out of 3M, serving clinical and hospital systems worldwide.
Major Telecomm, Broadband and Media provider in UK and Ireland.
Wayfair Inc. is a multi-billion dollar American e-commerce company based in Boston that sells furniture and home goods online.
Healthcare and Compliance startup with AI and IoT based solutions for B2B clients.
Adobe Systems, is an American multinational computer software company specialized in wide array of application for enterprises.
Google, multinational corporation and technology company focusing on online advertising, search engine, cloud, computer software and AI.
End-to-end AI/ML projects covering Agentic AI, RAG, deep learning, MLOps, and cloud deployment. Each project includes architecture diagrams, modular code, and CI/CD pipelines.
Agentic chatbot with specialised AI agents (router, research, analysis) orchestrated via LangGraph with real-time Tavily web search. Deployed serverless on AWS Fargate.
LangGraph LangChain Tavily AWS Fargate Jenkins Docker
AI-powered product recommendation chatbot using LangChain agents. Deployed on Kubernetes with Prometheus + Grafana production observability.
LangChain Kubernetes Prometheus Grafana Docker Flask
Medical Q&A chatbot powered by Retrieval-Augmented Generation. FAISS vector search over medical documents for grounded, source-cited answers.
RAG FAISS LangChain Flask Jenkins Docker
9-stage MLOps pipeline for phishing detection. MongoDB ETL, model factory (RF/XGB/SVM), MLflow on DagsHub, FastAPI serving, GitHub Actions CI/CD to AWS.
MLflow MongoDB AWS EC2 ECR GitHub Actions Docker
Full ML pipeline with MongoDB, SMOTEENN, model factory (5 classifiers + GridSearchCV), S3 model registry with F1-based promotion gate, Flask + AWS deployment.
scikit-learn XGBoost MongoDB AWS S3 EC2 GitHub Actions
Deep learning pipeline classifying adenocarcinoma from CT scans using VGG16 transfer learning. Full MLOps with MLflow, DVC pipelines, and AWS deployment.
VGG16 TensorFlow MLflow DVC AWS ECR GitHub Actions
End-to-end recommender with collaborative filtering (KNN, SVD), content-based (TF-IDF), MLflow tracking, Jenkins CI/CD, and GCP Cloud Run serverless deployment.
scikit-learn Surprise MLflow GCP Cloud Run Jenkins Docker
Production-style modular ML codebase predicting wine quality with ElasticNet regression. MLflow experiment tracking on DagsHub with Flask web serving.
ElasticNet MLflow DagsHub Flask Modular Code
Clean MLOps template demonstrating DVC for pipeline orchestration + data versioning and MLflow for experiment tracking. 3-stage reproducible pipeline.
DVC MLflow DagsHub scikit-learn MLOps Template
Notes on what actually changes when models leave the notebook and meet real users.
The model is rarely the hard part. What changes once real users depend on it — data contracts, evals, monitoring, cost ceilings, fallbacks and compliance — and why the estimate is usually three to five times the notebook.
Read on MediumFour ops disciplines, four different problems, and the question teams keep asking the wrong way round. How to work out which one you actually need before buying tooling for all of them.
Read on MediumBelow are the details to reach out to me!