Mohamed CHARRAKI

Senior data scientist driving industrial AI products that combine physics-based process models, machine learning and local LLMs. Work spans the full production stack – from signal acquisition on the plant floor to no-code ML components shipped as first-class citizens in the Industream AloT platform (FlowMaker, DataBridge, AI Engine). EAF Optimization — large-scale Electric Arc Furnace at a European steel plant – Hybrid physics + ML endpoint control: designed and deployed a hybrid steel temperature and chemistry prediction system combining a first-principles energy / mass balance model, an online offset learner and an ML residual corrector on Celox temperature, carbon (%C) and oxygen (ppm) measurements — cutting model error vs. pure physics and enabling closed-loop endpoint recommendations. – Measurable plant impact: contributed to an EAF optimization program delivering up to ~20% tap-to-tap time reduction and ~8% energy savings (kWh/t) at a major European steel producer. – Real-time operator cockpits in Grafana: built "Refining Phase Live Cockpit" and "Physical Model Comparison" dashboards (mixed InfluxDB + PostgreSQL data sources, Volkov Labs ECharts panels) with phase bands, power-off shading and heat-level annotations — adopted by process engineers for day-to-day heat analysis. – Production data pipeline from scratch: engineered reliable ingestion and feature pipelines from Siemens S7 / OPC UA signals into InfluxDB (time-series) and PostgreSQL (heat-level KPIs), with correct timezone handling, power-on / power-off logic, basket-level charging analysis and steel-grade joins used by every downstream model and dashboard. – Anomaly detection and drift monitoring: shipped input-validation and distribution-shift detection layers that surface outliers and frozen sensors before they poison model predictions; designed to be marketable as a standalone product for plant quality teams. – Causal ML for endpoint control: applied PCMCI+, LINGAM, ICP and DoWhy to EAF signals to move from black-box correlations to defensible cause-and-effect recommendations for process engineers. Industream Platform - Flowmaker AloT (no-code / low-code for industry) – Owned the data-science component library: designed and built reusable FlowMaker boxes for statistical preprocessing, signal processing and ML on time-series, image and sound data — so that domain experts and junior engineers can assemble pipelines visually (Node-RED-style) instead of writing Python. – Published an exhaustive preprocessing catalog across three modalities: time-series (scaling, transformations, decomposition, STL streaming), image (pooling, striding, batch norm, dilated convolution) and sound (STFT, MFCC, Mel spectrogram, VAD) - serving as the reference for internal and customer-facing development. – Forecastream probabilistic forecasting module: shipped a zero-training, drop-in forecasting box returning p10 / p50 / p90 quantile bands on any plant signal; deployed live on 12+ EAF cooling panels and packaged as a first-class Industream product ("Forecastream - AI Engine"). – Minutes, not months: cut time-to-deploy for new ML use cases from weeks of custom code to a few FlowMaker boxes on a canvas, supporting the platform promise of "plant signal to AI decision in hours, not months". – ONNX-first model packaging: standardized the workflow from Python training notebooks to ONNX artifacts served by the AI Engine – enabling identical inference on edge, on-premise and cloud runtimes with a single build. Hybrid Model + Local LLM — Plant Advisor (trip & breakdown prevention) – Architected a local-LLM advisory layer on top of hybrid physics + ML models: the LLM consumes live signal context, physics-model predictions and ML anomaly scores, then explains risk and proposes countermeasures to operators in natural language – all fully on-premise, no data leaves the plant. – Trip & breakdown prevention focus: targeted high-impact failure modes (cooling panel overheating, sensor drift, abnormal energy patterns) with early-warning reasoning blending deterministic physics constraints and learned residuals. – MCP-based orchestration: wired the LLM to plant tools (InfluxDB queries, model inference, Grafana links) via Model Context Protocol, so operator questions map to grounded, cited answers instead of hallucinated ones. Mentoring & production engineering excellence – Mentored junior data scientists through their first production deployments, with a strict focus on: – Training-serving parity — same preprocessing, same features, same library versions in notebook and production. – Leak-free feature engineering — no look-ahead, timezone-safe windows, validated on held-out periods. – Fast, reversible deployments — Dockerized services, versioned models, canary rollouts, easy rollback. – Observability by default — every model emits input stats, prediction distribution and health metrics to Grafana. – Code review discipline — pull requests, pytest suites, type hints and reproducible notebooks before anything ships.

Developed, trained and deployed four generations of deep learning models to control a large-scale Electric Arc Furnace, directly influencing voltage control on large-scale machinery — preventing equipment damage, enhancing safety, reducing downtime and improving production efficiency. – Closed-loop voltage control with deep learning: shipped successive generations of neural-network controllers acting on real plant signals; received an official SMS group reference letter highlighting "exemplary" performance, outstanding perception of emerging problems and reliable results. – Quantified production impact: AI-driven voltage control deployed on two plant shells increased power-on time by ~10% and added ~2 additional heats per day — translating to tens of thousands of additional tonnes of liquid steel per year and a seven-figure annual margin uplift per furnace for the customer. – Predicted overheating events with high accuracy, enabling real-time power adjustments that avoided equipment damage, protected operational stability and reduced unplanned downtime. – Explainable AI for process control: designed XAI components that turned model predictions into actionable countermeasures for operators, improving decision-making and control-room trust. – Feature engineering grounded in domain physics: collaborated with process engineers to craft highly predictive variables (energy history, hot heel, melting / refining windows, O2 volume) that materially improved model quality over naive ML baselines. – Anomaly detection on plant sensors: built autoencoder-based detectors that flagged faulty sensors to customers before they corrupted both process KPIs and model accuracy. – Customer-facing rollouts: monitored live model performance, gathered operator feedback and iterated models to match evolving operational goals. – Interactive dashboards: delivered Plotly / Streamlit dashboards and Grafana cockpits surfacing the most influential signals affecting EAF performance. – Engineering discipline: wrote pytest-covered code, reviewed teammates' pull requests, and shared reusable Python modules on GitHub — contributing to a healthy engineering culture in the Data Science team. – LLM applications: prototyped Large Language Model use cases and prompt engineering for internal tooling.

– Predictive maintenance in steel: applied machine learning and deep learning time-series models to predict anode temperatures in Electric Arc Furnaces, improving reliability and reducing downtime. – Data quality first: implemented data validation pipelines to ensure robust model training on noisy industrial data. – Feature importance and anomaly detection: benchmarked advanced feature-importance algorithms and AI-based anomaly detectors to prevent failures before they occurred. – Engineering rigor: used pytest for unit testing and Plotly / Streamlit for real-time model monitoring dashboards.