GenAI Expert

4/2023 – 7/2023

Fine tuning of an LLM (meta/Llama_2) to improve customer queries and Q&A. Respond more consistently, learning focus, reduce hallucination, gain knowledge on insurance data. Vector averaging of word embeddings was used to measure model performance.

Llama_2 | Hugging Face | Pytorch | Embedding Distance | LAMINI | FastAPI | Pandas | Numpy

Natural Language Processing, Python, Pytorch

2/2023 – 3/2023

Summary review on LLM concepts, including private and open-source models, comparison of model building architectures, various MT benchmarks (HellaSwag, HumanEval), various LLM studios (H20, LAMINI) offering such services, and ethics and trustworthiness aspects of it.

Natural Language Processing, Python, Pytorch, Spracherkennung, Text Mining, Textklassifikation

6/2022 – 12/2022

Automated Scientific Discovery – Use of ML methods to ease the cycle of scientific discovery. Implemented an automated ML platform with several explainable AI and causality frameworks.
Tools:
Dask | Ray | Pytorch | Python | Optuna | LIME| SHAP | Anchor | ASV | SHAPLEY FLOW | Captum | FastAPI

Cloud (allg.), Python

3/2022 – 6/2022

Development of a feature-store to help boost model development process. Automation of machine learning models using the complete MLOps stack of AWS including CI/CD, model pipelining, deployment, and scheduling.
ETL scripts to load data google analytics, flatten the key/value structure of the data, and finally save it in AWS S3. Built a model for text labeling to better understand student feedbacks.

Apache Spark, Business Intelligence (BI), Datenanalyse, ETL, Keras, Natural Language Processing, Postgresql, Python, Pytorch, Tensorflow, Textklassifikation

3/2021 – 4/2021

Implemented a Twitter bot which search through tweets on topics of interest, summarizes the text on those tweets, perform sentiment analysis and then publish the story on the timeline. Found @empirisch2. Tools used were Python| Tweepy |NLTK | Hugging Face T5 | Sentiment Analysis

Keras, Natural Language Processing, Python, Tensorflow

2/2021 – 2/2021

Implemented a search engine for DIY projects. The platform is live at findingdiydotcom. Tools used were Python| Django | Selenium | Hugging Face T5 | NLTK | Sentiment Analysis.

Django, Natural Language Processing, Selenium

1/2021 – 1/2021

Implemented a search engine for 3D printable models. The platform is live at 3dfindabledotcom. Tools used are Python | Django| Selenium | VGG16.

Django, Natural Language Processing, Selenium

3/2019 – 8/2019

STOCK PRICE PROJECTIONS
Price forecasting of blue-chip stocks.
Python | SQL | LSTM | RNN | Boosting | SVM | GNU | Scikit-Learn | Keras | FBProphet | Optuna | State-Space Modeling | ARIMA | Kalman Filter | ta-lib

Apache Hadoop, Gradient Boosting, Keras, Natural Language Processing, Python, Rekurrentes Neuronales Netzwerk (RNN), Scikit-learn, Support Vector Machine

4/2017 – 12/2021

MODELS FOR CROSS-SELLING (02.2021 - 11.2021)
Several machine learning models are built to help facilitate the cross-selling activities of a telecom operator. The models are built to, 1) identify the likely customer for cross-selling, 2) the most suitable product for that customer and lastly, 3) the right channel to approach that customer. Campaign success rate increased by 17%.
Python | SQL | DWH | CI/CD | Oozi | Tableau | Google Analytics | Big Query | Boosting | Bagging | Optuna


MODELS FOR UP-SELLING (03.2021 - 08.2021)
Built machine learning models to help facilitate the up-selling activities. This is to facilitate the base-management challenges, i.e., increase ARPU per customer by up-scaling their tariffs. Campaign success rate increased by 28%.
Python | SQL | DWH | CI/CD | Oozi | Tableau | Google Analytics | Big Query | Scikit-Learn | Optuna

LIFT AND SHIFT 08.2020 - 01.2021
The end-to-end execution of household identification use-case was too time consuming and was block other pipelines. Therefore, a lift and shift approach was adopted to fully migrate the use-case to the GCP, including its ETL.
Pyspark | Cloud Storage | GCP Dataproc | GCP Data Studio | GCP Cloud Composer | AirFlow

HOUSEHOLD IDENTIFICATION (04.2020 - 01.2021)
Built a model to identify customers living in same household for a telecom operator. Extensive explorative data analysis is performed to identify potential levers to help improve the matching criteria.
Pyspark | Hive | SQL | CI/CD | Oozi | Tableau | Google Analytics | Big Query | MLIB | Graph Theory


PROACTIVE MAINTENANCE (07.2019 - 03.2020)
Proactively detect the root causes of network problems. A thorough time-series analysis is performed and both induction (ML) as well as deduction-based models are built. Time to resolve the ticket was reduced by half.
Pyspark | Hive| SQL | Power BI | CI/CD | Oozi | Tableau | ElasticSearch


TARIFF OPTIMIZATION (02.2019 - 05.2019)
Sketch a relation between the old products and current customer base and identify any gaps in the current portfolio which hinders the operators in increasing their customer base. Additionally, analyze the need for any shadow products for retention and customer base management.
R | R FOR OPERATIONS RESEARCH | Google ortools | Crone | R&D| Clustering


VALUE BASED NETWORK ROLL-OUT (05.2018 - 02.2019)
Based on several customer satisfaction KPIs, a model is built to efficiently target network rollout and upgrade activities by investing in sites with poor network quality scores. This results in saving of several million euros as network rollout is a high budget activity i.e., a small % improvement in investment strategy results in considerable savings.
Pyspark | Hive | Power BI | Traffic Forecasting | xgboost | Oozi | R&D | Analytical Models

CUSTOMER EXPERIENCE (09.2017 - 05.2018)
Several models are built to measure customer experience on various service aspects. E.g., customer satisfaction via NPS, bill shock, customer experience w.r.t network quality, as well as service line interactions. Several inductions (ML) and deduction-based models are built.
Pyspark | Hive | Python | SQL | Power BI |Oozi | Analytical Models


CHURN & RETENTION (05.2017 - 09.2017)
Built a model to predict customer churn. Additionally, a detailed analysis on churn reasons is performed by calculating the SHAPLEY values. After A/B testing, an ultimate 17% reduction in churn was achieved.
Python | SQL| Power BI | DNN | Boosting | SVM | Bayesian Statistics | SHAPLEY | Crone Job | Scikit-Learn | Keras


TRAFFIC LOAD FORECASTING (04.2018 - 07.2020)
Forecasting network traffic per cell basis. This project was the pre-requisite of value-based roll-out.
Python | SQL| LSTM| RNN | Boosting | SVM | GNU | Scikit-Learn | Keras | FBProphet | Optuna | State-Space Modeling | ARIMA | Kalman Filter | DeepAR | AWS SageMaker

Apache Hadoop, Apache Spark, Big Data, Elasticsearch, Google Analytics, Maschinelles Lernen, Power Bi, Python, R (Programmiersprache), Scikit-learn, Tableau

2/2016 – 2/2016

Worked as an application to develop battery softwares

C, C++, Python