Projects

How I helped a company adopt Claude organization-wide: privacy and compliance groundwork, connecting internal systems from one-click connectors to custom MCP servers and browser flows, and training every team to use it.

We helped FileAI adopt Generative AI for document processing at scale: LLM and VLM models that lifted OCR and form-filling accuracy on complex multi-language documents, deployed on AWS.

A Terraform module for a SageMaker Studio domain with no internet egress: VPC-only networking, Identity Center authentication, a customer-managed encrypted EFS, and per-user POSIX-isolated storage.

I built an AI website support chatbot for Tellofy: indexing site content and knowledge bases to answer customer questions directly, deployed on AWS with Bedrock.

An agent that transcribes recorded conversations, embeds them, makes them full-text and semantically searchable, and layers a chat agent on top that answers with timestamped references back to the source audio.

An internal project we built to evaluate, deploy, and benchmark self-hosted embedding models against managed ones. Private embeddings are cheap to serve and cheap to reindex, and our internal benchmark confirmed them as a viable foundation for fully private AI.

I partnered with ComeToTheClouds to bring Generative AI to their AWS environment: a Retrieval-Augmented Generation solution on AWS Bedrock.

Core engineer for the AWS landing zone and on-prem migration of one of the first hospitals in Germany to run its infrastructure entirely on AWS: delivered in pure Terraform (VPC, EC2, RDS, S3, monitoring, encryption). Featured as an official AWS case study.

A multi-account, governed data mesh on AWS for a leading global reinsurer: producer, governance, and consumer accounts with Lake Formation tag-based access control, delivered in Terraform, cutting data-access provisioning from weeks to minutes.

A private, self-hosted speech-to-text engine I built with Whisper, optimized for high-throughput inference with speaker diarization. A queue-driven GPU pipeline on AWS that autoscales to zero when idle, so audio never leaves your account and idle cost is nothing.

Master thesis research published in Image and Vision Computing (2024): extending behavioral features with authentication phrase information.

Generative-AI quality control for call-center recordings: speech recognition, speaker diarization, and retrieval-augmented LLMs that answer questions about a call and grade it against configurable criteria.

Automatic parking slot detection and occupancy classification from camera imagery, published in Expert Systems with Applications (2023).

Winning solution (1942/2000) to FERIT Osijek 2022 RPSLS competition: an LSTM that predicts the opponent robot next move from the move histories of both players.

A customizable, natural-sounding text-to-speech model for customer support: voice-cloned by fine-tuning Tacotron2 on a proprietary dataset collected with Prodigy, with prosody realistic down to the in-breath between sentences.

Multimodal emotion recognition on the MELD dataset, built at the PSI:ML7 machine learning institute: fusing text and audio to recognize emotion in conversation.

Full implementation of the original Transformer paper (Vaswani et al., 2017) in PyTorch, written as learning material.

A variety of sequence model architectures implemented from scratch, structured to serve as learning material.

Implementation of the neural style transfer paper: transferring artistic style between images with deep networks.

End-to-end classical ML project: data collection, feature engineering, and price prediction for real-estate listings.