Jul 2026 — Present
AI Lead
BASF · Ludwigshafen, Germany
Leading AI development for enterprise document intelligence and multi-agent solutions connected to real business workflows.
Research Scientist · AI Lead
Advancing tabular learning and building enterprise AI.
My research focuses on tabular deep learning and numerical feature representations. Alongside this, I lead the development of production-grade GenAI and agentic systems.
Input features
Spline transforms
Learned representation
Learnable parameters
Tabular Deep Learning
Neural architectures that learn effectively from heterogeneous, structured data.
Feature Representations
Numerical encodings, spline bases, and adaptive transformations for better learning.
Agentic AI
Reliable AI systems that combine reasoning, tools, and production-grade workflows.
Research focus
Representation is part of the model.
My research asks a practical question: how should numerical and categorical features be represented before a neural network starts learning?
I study tabular architectures, spline-based numerical encodings, target-aware and learnable representations, and the interaction between preprocessing and model design. Alongside this work, I develop applied generative-AI and agentic systems for real-world workflows.
Tabular DLNumerical encodingsSplinesFeature learningAgentic systems
From Uniform to Learned Knots: A Study of Spline-Based Numerical Encodings for Tabular Deep Learning
A systematic study of B-, M-, and I-spline encodings with uniform, quantile-based, target-aware, and learnable knot placement.
Abstract
Numerical feature representation is a central design choice in tabular deep learning. This work systematically compares spline-based representations across spline families, knot placement strategies, representation sizes, and neural backbones. The study examines uniform, quantile-based, target-aware, and gradient-learned knots across classification and regression tasks, and derives practical guidance for choosing expressive yet efficient numerical encodings.
B-splinesI-splinesLearnable knotsPLE
Read on arXiv Mambular: A Sequential Model for Tabular Deep Learning
An adaptation of the Mamba architecture for regression, classification, and distributional regression on tabular data.
Abstract
Mambular adapts selective state-space models to tabular data. It provides a flexible framework for classification, regression, and distributional regression while exploring how sequential modelling can capture interactions among structured features.
MambaState-space modelsDistributional regression
Read on arXiv Applied AI systems
Research thinking, deployed at scale.
Selected industrial work described at a public, architecture-focused level.
01Document intelligence · GenAI
Enterprise Document Extraction
An LLM-powered OCR platform that extracts business documents, maps required SAP data, and supports order submission.
Document intelligence · GenAI
Enterprise Document Extraction
An LLM-powered OCR platform that extracts business documents, maps required SAP data, and supports order submission.
- OCR, document understanding, and structured LLM extraction
- SAP data mapping and validation across document types
- Asynchronous processing, observability, and scalable deployment
Azure OpenAIDocument IntelligenceFastAPIMongoDBRabbitMQCeleryMLflowKubernetes
02Multi-agent systems · Enterprise AI
Enterprise Multi-Agent Assistant
A multi-agent assistant integrated with web portals, Microsoft Teams, Salesforce, and internal enterprise services.
Multi-agent systems · Enterprise AI
Enterprise Multi-Agent Assistant
A multi-agent assistant integrated with web portals, Microsoft Teams, Salesforce, and internal enterprise services.
- Connects agents, tools, and services owned by different teams
- Supports product, order, and customer-service workflows
- Retrieves product data and documents and performs authorized actions
LLMsMulti-Agent SystemsMicrosoft TeamsSalesforceEnterprise APIsAccess Control
03Time series · AutoML
Commodity Price Forecasting
A production forecasting platform for commodity-price intelligence, combining reusable modelling workflows with cloud deployment.
Time series · AutoML
Commodity Price Forecasting
A production forecasting platform for commodity-price intelligence, combining reusable modelling workflows with cloud deployment.
- Time-series feature engineering and forecasting model development
- Automated model selection, evaluation, and reproducible experimentation
- Cloud deployment and integration with business decision workflows
PythonTime-Series ForecastingDeep LearningAutoMLMicrosoft AzureMLOps
04Predictive analytics · Healthcare
Healthcare Claims Analytics
Predictive modelling and analytics infrastructure for understanding and managing healthcare-claim denials.
Predictive analytics · Healthcare
Healthcare Claims Analytics
Predictive modelling and analytics infrastructure for understanding and managing healthcare-claim denials.
- Machine-learning models for claim-denial prediction
- Large-scale data processing and analytical workflows
- Integrated analytics and visualization platform
Machine LearningBig DataPredictive AnalyticsVisualization
Open source
Research translated into tools.
APythonDeepTab
A unified, scikit-learn-style library for modern tabular deep learning models, evolved from the Mambular project.
DeepTab brings modern neural architectures for structured data behind a consistent estimator interface. The project focuses on reproducible training, modular preprocessing, stable configuration, and practical model comparison.
- Unified APIs for multiple tabular deep-learning architectures
- Classification, regression, and distributional modelling workflows
- Designed for experimentation as well as production-oriented use
BPythonPreTab
An extensible toolkit for numerical feature representation, basis expansion, and preprocessing for tabular machine learning.
PreTab separates feature representation from the downstream model, making numerical encodings easier to study, compare, and reuse. Its direction includes adaptive and data-aware transformations alongside a broader family of basis expansions.
- Composable numerical encodings and spline families
- Extensible basis-construction and feature-placement framework
- Scikit-learn-compatible transformations for research and practice
Career & research journey
Building production AI, advancing structured-data research.
A professional track paired with an ongoing academic research track.
Jun 2025 — Jul 2026
Principal Data Scientist
BASF · Ludwigshafen, Germany
Solution architecture for multilingual GenAI extraction systems, with a focus on reliable LLM workflows and production AI.
Oct 2023 — Present
PhD Researcher in Artificial Intelligence
TU Clausthal · Parallel research track
Tabular deep learning, numerical feature representations, spline-based methods, and time-series forecasting.
Oct 2020 — Jun 2025
Senior Data Scientist
BASF · Ludwigshafen, Germany
Production AI systems spanning forecasting, deep learning, AutoML, MLOps, and AI product development.
2017 — 2019
M.Sc. Cognitive Science
Universität Osnabrück · Grade 1.3
Specialization in Artificial Intelligence and Neuroinformatics.
Mar — Jun 2019
Data Scientist Intern
BASF · Ludwigshafen, Germany
Commodity-price forecasting models deployed on Microsoft Azure.
Sep 2015 — Sep 2017
Consultant → Data Scientist
KPMG India · Bengaluru, India
Machine learning for credit risk, résumé scoring, healthcare claims, and large-scale analytics.
Jul 2011 — Sep 2015
Associate → Senior Software Engineer
Accenture · Bengaluru, India
Forecasting, customer analytics, industrial sensor processing, data warehousing, and ETL systems.
About
Research between theory and production.
I am a Research Scientist and AI Lead at BASF and a PhD candidate at Clausthal University of Technology. My work sits at the intersection of machine learning methodology, open-source research software, and production AI.
I care about methods that are not only accurate, but also reproducible, computationally grounded, and genuinely useful to practitioners.
Contact
Let’s discuss research, open source, or applied AI.
For research collaborations, technical discussions, or open-source work, send me a note. The form opens your email app with the message ready to send.