Marcus Lehmann verfügbar

Marcus Lehmann

Machine Learning Engineer & Data Visualization

Profilbild von Marcus Lehmann Machine Learning Engineer & Data Visualization aus Weinstadt
  • 71384 Weinstadt Freelancer in
  • Abschluss: nicht angegeben
  • Stunden-/Tagessatz:
  • Sprachkenntnisse: deutsch (Muttersprache) | englisch (verhandlungssicher)
  • Letztes Update: 05.05.2021
Profilbild von Marcus Lehmann Machine Learning Engineer & Data Visualization aus Weinstadt
Hallo, ich bin Marcus - ein leidenschaftlicher Data Nerd & Full Stack Developer & Entrepreneur.

Seit meiner Schulzeit liebe ich die Entwicklung von Webanwendungen. Heute bin ich mehr denn je davon begeistert, wie Technologie Dinge ermöglicht, die vor weniger Zeit undenkbar waren.

Meine aktuellen Schwerpunkte:
- Fullstack Entwicklung in der Cloud (Amazon Web Services)
- Data Collection (Public Data, Spatial Data, Site Scraping, API scraping)
- Machine Learning (XGBoost, scikit-learn, PyTorch, Tensorflow)
- Interactive Data Visualization (d3.js)

Haben Sie spannende Projektideen und wollen diese schnell und kompetent umsetzen?
Dann kontaktieren Sie mich für ein unverbindliches Beratungsgespräch.

Mehr als 5 Jahre Erfahrung:
JavaScript, SQL, Software Development, Data Analysis, Docker, AWS, Geospatial Analysis, REST API, HTML5, CSS, RESTful Services, Fullstack, Git, JSON, Java, Microservices, Backend, GitHub, HTML, Linux, Data Visualization, XML, d3.js

1-5 Jahre Erfahrung:
NumPy, Scikit-Learn, Vue.js, TensorFlow, NoSQL, Bootstrap, Python, Machine Learning, MySQL, xgboost, Pandas, Data Science, ReactJS, Scrum, PyCharm, Neural Networks, Data Engineering, Jupyter, PostgreSQL, Angular, IntelliJ, Node.js, PostGIS, Deep Learning, Leafet, Flask, jQuery, AngularJS, Computer Vision, Ansible, NPM, MongoDB, Apache Cassandra, Spark, Android, Geographic Information System, Natural Language Processing
  • 05/2020 - 07/2020

    • ReachNow
    • 250-500 Mitarbeiter
    • Telekommunikation
  • Map Matching of vehicle position data
  • Input: Billions of noisy vehicle telemetry data points (position, direction etc.)
    Output: Vehicle trajectories matched to street
    The algorithm is scheduled periodically on AWS Batch to process historic telemetry data stored on S3. I used GraphHopper Map Matching and implemented a Particle Filter that leads to matched trajectories of high quality. The quality of the processed trajectories is measured and can be monitored using AWS CloudWatch.
    AWS Batch, Lambda, S3, CloudWatch, Docker, Python, Pandas, OSM, Graphhopper

  • 01/2020 - 04/2020

    • ReachNow
    • 250-500 Mitarbeiter
    • Transport und Logistik
  • Frontend and Backend for managing models / clusters for a Fleet ML service
  • A UI (and Backend for the Frontend) was built to make an ML solution easy to expand to new cities and easy maintanable. The UI makes it possible to manage and extend the services for people that are not deeply involved into the technical details. A specialty about the project is the storage of the meta data (cities, clusters, POIs) via GitHub API to allow the 4-eyes principle via GitHub PRs and enable versioning and rollbacks.
    AWS ECS, S3, Docker, GitHub API, Python, ol3js, Vue.js, tailwindcss

  • 09/2019 - 12/2019

    • ReachNow
    • 250-500 Mitarbeiter
    • Transport und Logistik
  • Interactive Vehicle ETA prediction visualization
  • For a shuttle service 3 different estimations of arrival time (ETA) are used. To track their performance over time and by the given circumstances these ETAs were collected and compared by their actual arrival time. An interactive visualization (Backend & Frontend) was created to analyze the different ETAs.
    data collection: AWS Lambda, S3
    backend: AWS ECS, Docker, Python Flask
    frontend: Vue.js, ol3js, d3.js

  • 05/2018 - 12/2018

    • moovel Group GmbH
    • 250-500 Mitarbeiter
    • Transport und Logistik
  • Fleet redistribution service
  • Based on the predicted demand a shuttle service should redistribute idling vehicles to areas with more demand to raise the chance of getting requests from customers. I built a service that exposes a REST API to get redistributions based on the current predicted demand, a set of vehicles and the area the vehicles are operating in. Based on given costs (duration the vehicle needs to the redistribution point) and the gain (covered demand) the most optimal solution will be returned.

  • 06/2017 - 04/2018

    • moovel Group GmbH
    • 250-500 Mitarbeiter
    • Transport und Logistik
  • Demand Prediction and Visualization
  • To build an effective on-demand service it is good to know the demand in advance to react accordingly. I built a demand prediction service. Work done: Scraping (current Events, Weather data), Data cleaning & Feature Engineering, REST endpoint for building hex-bins, Model Training, Accuracy monitoring, Automating everything on AWS Lambda & Batch, REST inference endpoint, Interactive Visualization.

  • 05/2016 - 05/2017

    • moovel Group GmbH
    • 250-500 Mitarbeiter
    • Transport und Logistik
  • Portland Bus Delay Prediction and Visualization
  • A big problem for many people and modern urban mobility are delays of busses. The goal of this project was the accurate prediction of the arrival of busses at their destinations. The GTFS realtime data feed was provided by TriMeT, a bus operator in Portland USA. The positions of the busses have been collected and used to train a predictive model of the deviation of the arrival time of the busses.

  • 06/2012 - 05/2015

    • < 10 Mitarbeiter
    • Internet und Informationstechnologie
  • CTO & Software Engineer
  • ROOMAPS was an innovative indoor navigation & information system for smartphones with individual map-design and modern technology for positioning indoors.

    The core of the ROOMAPS technology consisted of:
    Map Import & Rendering. Import map features based on AutoCAD files and store as PostGIS geometries. The tile server renders the tiles on-the-fly when requested. Served through AWS CloudFront they are cached for the next requests. For rendering the Java Topology Suite and GeoTools have been used.

    Navigation Mesh Building & Routing. Based on the walkable area a navigation mesh (a graph which consist of convex connected Polygons) is created automatically. This is used for the A* routing algorithm to find shortest pathes.

    Indoorpositioning. An extensible approach of a particle filter and Map Matching based on Navigation Meshs was used to get the positioning of a smartphone. Depending on the environment all the available information a smartphone can gather was fed into the algorithm to improve its accuracy: Acceleration, Gyro, Magnetic Field, Bluetooth, WiFi, Barometers.