Muhammad Zohaib S.

I build models that run in production on messy real-world data, and I know when they are wrong. My current models handle real-time traffic and audience prediction across more than 15,000 screen locations in 24 countries, spanning Europe, Asia and South America. PhD from NTNU Norway in probabilistic deep learning and signal processing. 476 citations, first-author publications in Mechanical Systems and Signal Processing, Measurement and Engineering Structures. But what matters is what has shipped. 📊 Commercial analytics at scale As Senior Data Scientist for a Danish out-of-home advertising analytics company: Real-time traffic prediction across 15,000+ screen locations in 24 countries, on three continents Footfall and audience modelling, turning raw movement data into validated audience numbers advertisers are billed against Demographic prediction, inferring audience composition from behavioural and contextual signals Campaign validation across markets, reconciling modelled reach against measured ground truth A production text-to-SQL assistant I built and deployed: plain-language questions converted to validated queries across InfluxDB and ClickHouse, now serving 300+ daily users across several countries ⚙️ Industrial and sensor data A condition monitoring system running live on the Norwegian national rail network: 10 trains, 100 distinct signal types, continuous production use. It identified 20+ wheel-slip hotspots that went to infrastructure teams as targeted intervention points. A deployed RAG assistant over technical documentation with source-cited answers. FAISS, SentenceTransformer embeddings, containerised and in daily use. An open benchmark dataset for structural damage detection, co-developed with Stanford, UNSW and University College Dublin What ties these together Both are the same problem. Large volumes of noisy operational data, no clean labels, and a decision at the end that somebody is accountable for. Time-series modelling and forecasting. Anomaly detection without labelled failures, which is almost always the real situation. Feature engineering on raw signals. Spectral and envelope analysis. Predictive maintenance and fault diagnosis. Geospatial and sensor data pipelines. And getting all of it into production rather than leaving it in a notebook. The part most people skip My research specialism is uncertainty quantification: calibration, out-of-distribution detection, and building systems that recognise when they are outside their reliable range and escalate rather than guessing confidently at nonsense. Whether it is a fault diagnosis an engineer will act on or an audience number a client is billed against, the difference between a demo and a product is knowing when the model should not be trusted. Most ML work skips this. It is what I published on for four years. LLM and AI tooling RAG pipelines over technical and operational data. Text-to-SQL over warehouses. Fine-tuning open models with LoRA where a hosted API is not an option. Evaluation and guardrails so generated output is checked before it reaches anyone. How I work Working MVP in 30 days for a well-defined problem. If your problem is not well-defined yet, I will say so, and help you define it before we build anything. I would rather lose a week scoping than deliver something technically correct that solves the wrong question. Stack Python, PyTorch, TensorFlow, scikit-learn, NumPy, SciPy, Pandas SQL, InfluxDB, ClickHouse, Snowflake Azure ML, MLflow, Docker, Git, Power BI MATLAB, C/C++ COMSOL, ABAQUS for physics-based modelling Getting started Send me your problem, and a sample of your data if you have one. Within a day I will tell you whether it is solvable, roughly what it would take, and whether I am the right person for it. If I am not, I will say so and point you at what you actually need.