Five industrial AI systems — delivered end-to-end, from requirements to deployment.
📦 Logistics · Cost Intelligence
Logistics Cost Estimation Platform
Problem
Cost estimations for logistics planning were calculated manually across spreadsheets — a multi-day process prone to inconsistency and version drift.
Approach
Built a modular Streamlit platform integrating packaging, transport, warehouse, customs, environmental, and financial modules into a single, validated workflow with persistent JSON storage and exportable reports.
Python
Streamlit
JSON storage
Modular architecture
↓ 90% processing time — days → ~40 minutes
View on GitHub →
⚑ Enterprise-safe: abstracted business logic; no proprietary rates or client data.
🤖 AGV · Real-Time Fleet
AGV RTLS Dashboard
Problem
No live visibility into AGV fleet positions, zone utilisation, or anomaly patterns inside a production facility.
Approach
Rebuilt a Python/Flask tracking system with MQTT ingestion at 3 Hz, WGS84 → UTM → affine coordinate transforms, MySQL persistence, and a Streamlit dashboard rendering live trajectories, heatmaps, and zone overlays. Docker-orchestrated broker, database, and UI services.
MQTT
MySQL
Docker
Streamlit
Node-RED
Coordinate transforms
⚡ 3 Hz real-time ingestion · Full fleet visibility
View on GitHub →
⚑ Enterprise-safe: plant layout abstracted; no proprietary map data included.
📈 Analytics · KPI Intelligence
RTLS Trajectory Analytics Dashboard
Problem
Historical AGV position data in Snowflake lacked structured KPI reporting — speed, distance, and utilisation metrics were not actionable.
Approach
Developed a multi-tab Streamlit application (Snowflake or CSV input) with data cleaning, smoothing, KPI computation, and optional SVG plant-layout overlay. Exportable outputs and dynamic column/date filters enable self-service analytics.
Snowflake
Streamlit
Python
Plotly
SVG overlay
📊 KPIs: distance · speed · utilisation · exports
View on GitHub →
🔮 ML · Demand Forecasting
Material Consumption Forecast Dashboard
Problem
Material consumption per machine was unpredictable, with zero-consumption days inflating forecast errors when modelled with a single approach.
Approach
Designed a two-stage ML pipeline: a logistic classifier to flag zero-consumption days, followed by an LSTM regression model for magnitude prediction. Data is pulled and merged from Snowflake; modular codebase (fetch, prepare, forecast, train) is extensible to additional machines.
Snowflake
Dash / Plotly
LSTM
Logistic Classifier
Python
🧠 2-stage pipeline: zero-day detection + magnitude LSTM
View on GitHub →
⚑ Enterprise-safe: model architecture shared; training data and machine identifiers abstracted.
🤖 GenAI · Mobility Intelligence
Schaeffler Mobility Insight Platform
Problem
Identifying and quantifying disruptive mobility technologies at scale required analyst time and lacked traceability, variance tracking, or structured evaluation.
Approach
Built a multi-agent GenAI web platform using GPT-4o, LangChain, and a Dockerized microservices layer. Agents coordinate data retrieval, analysis, and insight generation with RLHF. The system tracks confidence scores, variance, and technology readiness — with REST/WebSocket APIs, modular feature flags, and automated reporting cycles.
OpenAI GPT-4o
LangChain
Flask
React
Docker
MySQL
RLHF
WebSocket
✅ Evaluation · Traceability · Multi-LLM fallback · RLHF
View on GitHub →
⚑ Enterprise-safe: Master Thesis project; business context abstracted, architecture shared.