I design and build AI/ML models end to end, from early ideation to production systems serving thousands of requests. My work starts with understanding the business domain and designing for its specifics, then accounting for where probabilistic models fall short by adding deterministic guardrails and monitoring to keep performance reliable in production.
Alongside the engineering, I drive projects forward, work closely with collaborators, and guide interns and master's thesis students.
Outside of work, you'll usually find me playing badminton or tennis, or out hiking.
Industries
Textile RecyclingAutomotiveMaterial ScienceSensor SystemsRetailMediaEducation
Applications
Agentic WorkflowsAgentic ChatbotsOptical Character RecognitionDemand & Price ForecastingAutomated Textile SortingAnomaly DetectionRecommendation SystemsRisk AssessmentDocument Information Retrieval
Skills
Python
LangChain
LangGraph
vLLM
PyTorch
PySpark
Apache Kafka
Prefect
n8n
Databricks
Azure
Docker
GitHub CI/CD
TypeScript
PostgreSQL
MongoDB
SurrealDB
FAISS
Languages
GermanLimited Working
EnglishFull Professional
TamilNative
KannadaNative
HindiProfessional
Certifications
Neural Networks & Deep Learning
LLM apps with LangChain & OpenAI
Azure Machine Learning Fundamentals
Azure Data Factory
GDPR for Businesses & Individuals
Experience
Machine Learning Engineer
TURNS (Wagner+Thummerer GmbH)- Building AI for textile-to-textile recycling — automated sorting of garments by fibre, colour and quality.
AI Engineer — Detection of Hidden Objects
Balluff EMEA- Compared YOLO, RetinaNet and Vision Transformers for object detection on industrial 3D radar imaging, capturing complex spatial patterns for quality assurance.
- Applied FMCW radar signal-processing as a pre-processing step to improve model training and detection accuracy.
- Developed AI methods to identify anomalies and defects in industrial products, reducing errors and improving production reliability.
Data Science Intern — Future Lab for Additive Manufacturing & Engineering
German Aerospace Center (DLR)- Analysed the emissivity of 3D-printed polycarbonate across different shapes and geometries.
- Investigated the temperature dependence of polycarbonate emissivity using a reference object in a 150 °C heating chamber.
- Used Python for data analysis, visualisation and presenting results to the team.
AI Engineer — Research & Development
Schwarz Dienstleistungen- Built an image-similarity model (Siamese network with triplet loss) to track retail products from warehouse to shelf without retraining for new products.
- Built an anomaly-detection system for defects in printed text and deployed a containerised PaddleOCR pipeline on an NVIDIA Jetson Nano.
- Created Debian packages enabling GPU access on the Jetson Nano from within Docker containers, and benchmarked edge inference latency.
Data Science Consultant
CoffeeBeans Consulting- Implemented a Named Entity Recognition model with a prior probability distribution, improving click-through rate by 8%.
- Designed and ran an A/B-testing strategy that reached statistical significance to validate improvements.
- Built a probabilistic push-notification recommender (1,000 notifications/sec) that lifted user interaction by 10%.
- Extended the SimDB semantic-search system with Faiss + Kafka for faster, higher-throughput retrieval, plus news grouping and sentiment-based ranking.
Computer Vision Intern
Facex.io- Developed an adaptive shadow-removal technique for outdoor video analytics under varying lighting conditions.
- Built an edge-optimised face-detection pipeline (HOG + SVM with random-forest landmark detection via dlib) at 90% accuracy.
K
Machine Learning Intern
Krishihub- Built a Seq2Seq demand-forecasting model using market trends and weather, reaching 90% accuracy for two major commodities.
- Built a lightweight LSTM price-forecasting model exposed as a REST API for a mobile app.
- Developed a remote-sensing solution extracting NDVI from satellite imagery to estimate harvest readiness.
i
Data Scientist
instinctus.ai- Built a super-resolution model that improved OCR and handwriting-recognition accuracy by 15%.
- Implemented OpenCV kernel-based image processing to extract survey-form data at ~60 ms per document.
- Built a Word2Vec contextual spelling-correction system, improving OCR accuracy by 10%.
- Built chatbots (Dialogflow, Wit.ai), a Flask + MongoDB annotation platform, and an AWS Lambda + S3 + SNS reporting pipeline.
Software Development Intern
Think201- Built e-commerce and education websites with PHP Laravel.
- Wrote Selenium and Appium automation scripts for mobile-application testing.
Software Development Intern
Hykmann Technologies- Built websites with PHP Laravel.
Industrial Projects
RoboTex
Fraunhofer IWKSAn automated robotic system that sorts used textiles for recycling — combining robotics with AI-powered image recognition and sensors to identify material composition and sort clothing automatically. Coordinated by Fraunhofer IWKS.
Project Partners
Fraunhofer IIS
Robotextile GmbHWagner+Thummerer GmbH
Siemens AG
Friedrich-Alexander-Universität Erlangen-NürnbergOpen Source
A lightweight, open-source vector similarity database for storing embeddings and running fast nearest-neighbour search. Built at CoffeeBeans Labs. I extended it with Faiss and Kafka improvements for higher-throughput retrieval.
In the Press
Texcare International · Messe Frankfurt · Aug 2025
How a recycling start-up is bringing Bosch, dm & Austrian Airlines into the textile loop
Hochschule Ansbach · 2023–2025
KI-gestützte Ermittlung qualitätsrelevanter Sortierkriterien für das Textilrecycling (DBU · TURNS)
ÖBB — Austrian Federal Railways · Mar 2025
„Zweites Leben für 1er-Panier": Schulterschluss bei textiler Kreislaufwirtschaft
dm-drogerie markt
Textilfaserrecycling — ein neuer Kreislauf für Alttextilien
Fränkische Landeszeitung (FLZ) · Sep 2024
Recycling und Inklusion: Diakoneo und Turns arbeiten in Bruckberg zusammen
Retail Optimiser · Mar 2026
Schwarz Group introduces new self-checkouts at Lidl and Kaufland
Awards
Best Startup — Digital Platform
EHI Retail Institute — Scientific AwardRecognised for our digital textile-circularity platform.
ViewINSPECT Award 2024
INSPECT Award — Machine Vision & AutomationHonoured for industrial inspection / detection work.
ViewBest Recommendation System
comScore — Vikatan.comRecommender behind Vikatan.com being ranked India's most engaged Tamil news website.
ViewEducation
M.Sc, Electrical Engineering — Smart Information Processing
University of StuttgartMaster's Thesis
Differential Privacy as a privacy-preserving technique in Federated Learning
Objective: Distributed training and privacy.
Supervisors: Prof. Michael Weyrich, Baran Can Gül
- Federated Learning decentralises deep-learning training across client devices — only local model weights are sent to a central server for aggregation into a global model, avoiding central data storage and easing server load.
- Exposed model weights are a security risk: an adversary can reconstruct a client's original data using generative models.
- Introduced Differential Privacy at training time via DP-SGD, which clips and adds noise to constrain each example's influence — trading some accuracy and convergence time for privacy.
- Used privacy accounting (epochs, noise, clipping bound) to quantify privacy, and searched for the optimal noise level that maximises privacy without degrading accuracy — evaluated on driver-profile identification across three client devices.