Senior Machine Learning Engineer building LLM & agentic systems, MLOps, time-series and computer vision that hold up in production.
6+ years taking ML and LLM systems in Python through the full lifecycle — data pipelines and feature stores, training and fine-tuning, evaluation and guardrails, deployment and monitoring on Azure and AWS — for clients in Norway, Sweden and Bangladesh. Now lead developer on Sensa AS’s industrial-AI platform at Cefalo.
Production ML and LLM systems — RAG and agents, anomaly detection, computer vision and the data platforms beneath them — most built end-to-end, from data pipeline to deployed, monitored model.
LLM assistant that lets industrial plant operators query live plant data in plain language: hybrid RAG over a Neo4j knowledge graph, InfluxDB time-series data and PGVector, with LangGraph agents, MCP tools and source-cited answers under data-governance rules to limit hallucinations.
End-to-end anomaly-detection platform for time-series sensor data: automated training with Isolation Forest and the pretrained Moirai v2 model, MLflow experiment tracking and model registry, Azure ML deployment and drift monitoring.
Trained and deployed a YOLOv11-nano object-detection model that detects and counts fish-feed pellets in real-time aquaculture video, enabling automated feed-rate optimisation for sensor-equipped fish farms.
Internal LLM résumé parser that turns PDF and DOCX CVs into structured, ATS-ready data with role and skill categories, plus a recruiter Q&A chatbot. Fine-tuned and benchmarked Llama 3.2, DeepSeek and Qwen with QLoRA (Unsloth); the locally hosted Llama 3.2 3B matched cloud-API quality at a fraction of the cost.
Data-orchestration and feature platform: Mage.ai ETL pipelines ingest high-throughput sensor and external data into time-series and feature stores, with reusable transforms shared by real-time anomaly detection and RAG retrieval.
SaaS travel chatbot for Resemolnets, grounded in a custom tour knowledge base, giving personalised tour recommendations and handling orders end to end. Owned prompt engineering, embedding strategy (PGVector) and safety guardrails from prototype to production.
Platform for Disruptive Ventures that tracks portfolio-company performance, sales and accounting. Third-party accounting and marketing data flows in through event-driven processing, with time-series sales forecasting in scikit-learn.
Kry’s digital primary-care platform, ingesting body measurements and lab blood-test results from third-party APIs, with predictive health-condition analysis in scikit-learn.
National ID OCR, face detection and billing microservices for Jamuna Bank’s eKYC platform — automating identity checks that were previously done by hand and cutting onboarding time across branches.
Lead developer for Sensa AS (Norway) and Cefalo’s internal AI products: RAG and agentic LLM systems, time-series anomaly detection, computer vision and the data and feature platform underneath. AI Task Force member and AI Hackathon mentor & judge. Ships production services used by 500+ users through an AI-agent-led workflow (Claude Code) with agent harnesses, guardrails, automated testing and end-to-end monitoring.
Tech lead for client products in Sweden — Resemolnets, Disruptive Ventures and Kry. Led cross-functional teams of up to 8 engineers across 6 client projects, owning system design, backend architecture, code review and delivery while staying hands-on in API and AI features.
Built Boarding Bay, Jamuna Bank’s eKYC platform: three production microservices — national ID OCR, face detection and billing — applying OpenCV, Tesseract OCR, PyTorch and TensorFlow to automate manual identity checks and cut onboarding time across branches.
Analysed mobile-app data for Romoni, an on-demand beauty-services marketplace, and built regression-based sales forecasts that informed pricing and marketing decisions.
Founded and ran an education initiative in Bangladesh that taught students technology and entrepreneurship — leading strategy, operations and program design.
Coursework: Artificial Intelligence, Pattern Recognition & Neural Networks, Probability & Statistics, Algorithms.
Predicted one-day match winners using Recursive Feature Elimination with Decision Tree, Random Forest and XGBoost models.