Working @ Amazon

Education @ RWTH Aachen

TEN Enigma @ Amazon

November 2021 - now

Role: Senior Software Development Engineer

since October 2024: spearheading a taskforce to optimize product titles (shortening & removing duplicated phrases) while ensuring no loss of searchability. - Project is in ideation phase

Leading the design and implementation of an SEO tool using LLMs to augment product content. The tool uses existing search information to identify product niches to reduce competition for products. Scaled the system to 1M products analzyed per day. Main challenges were to integration of LLMs as slow, expensive and probabilistic components - reduced costs by proposing and implementing pre-filters and scaled the LLM throughput 18x through parallelization. The project led to 9M+ augmented product titles between June and December 2024 with an average search impression increase of 56% covering 3P products on Amazon stores worldwide. Drove an experiment on 34k products with existing search issues (missing terminology / ranking issues). Augmenting these products using the SEO tool developed resulted in an 11% OPS uplift (~$625k). Used these insights to influence leadership to change the direction of the project for 2025 into focusing on these higher-demand products.

Led the design and implementation of a new system extracting search insights into actions (creating new products in stores, identifying product defects preventing them from being searchable). Ensured the extensibility of the system for new use-cases and supported the Business Team to identity new integration points with other teams to broaden the impact made by Enigma. Scaled the system to 50M+ search queries & 100M+ products / week resulting in 300M USD monthly revenue in 2024.

TEN Search Matching @ Amazon

January 2021 - November 2021

Role: Software Development Engineer II

Drove the development of an analysis tool to quantify indexing gaps and the impact of individual attributes in search.

Simplified the indexing logic for several attributes to be included by default for new store launches, resulting in higher search result relevance, increased sales, saving ~2 SDE weeks / marketplace launch.

A shortlist of my personal coding projects:

LLM Agent using tools

Technologies: Python, Apple MLX, LLama 3.2 models

Writing custom logic to call tools in an LLM Agent workflow. Extending the first trial using CrewAI and the LLama 3.2 3B model which failed to properly invoke the tools made available. Prompt-finetuning (multi-shot) helped resolve the issue.

Jupyter Notebook

Bike Sensors - Bluetooth Low Energy

Technologies: Golang, Bluetooth Low Energy Definitions, Javascript, Websockets

Integrating Bike Speed and Cadence Sensors by reading their values via Bluetooth Low Energy (BLE). Displaying the speed, cadence and calculated power output in real-time in a WebApp.

Github repo

Portfolio Tracker

Technologies: Python (Pandas & Streamlit), Cron

Simple Portfolio tracker as my broker does not offer an overview of the portfolio development over time. Retrieve asset prices (daily Cron job) & display total values.

Github repo

Plant Database Retrieval

Technologies: Python, Langchain, FAISS, LLM Text Embeddings

Building a plant retrieval system. The plant data is retrieved from Wikipedia, transformed into a dense-vector representation using the MiniLM-L12 LLM and stores in a FAISS vector-store for semantic similarity retrieval.

Jupyter Notebook

Virtual Meetings

Technologies: Golang, Socket IO, Typescript/ReactJS

(In-progress) Evaluating and implementing a real-time virtual meeting system: meeting as avatars in a 2D map, where you can interact with other users and objects in the meeting room.

Github repo

Generate your own Book

Technologies: Python, GPT-3.5, GPT-4, Dall-E 2, Typescript/ReactJS

Generate your own book after providing content detail. The GPT4 and GPT-3.5 LLMs write the text and Dall-E 2 creates illustrations matching the text. Supports 2 languages (German and English). Main challenges: how to ensure the LLM generates age-appropriate content and how to steer the Dall-E 2 model to generate consistent, relevant images: prompt engineering

Project Homepage

Kafka Event Streams

Technologies: Apache Kafka, Telegram Bots, Python, Golang, Typescript/NodeJS, YOLO object detection

Distributed monitoring system: sensors, monitors and actors. Sensors include reading temperature sensors, presence sensors, webcams. Monitors include person detection using YOLO. Actors include Telegram Bots to send alert notifications.

Github repo

Interested in knowing more?

In case you would like to know more about my previous projects, have a question on career development at Amazon, then drop a message via LinkedIn or my Instagram Profile. Alternatively you can check out my Github repositories for project documentation