--- title: "Tobias Sterbak — Freelance AI & Machine Learning Engineer" description: "Freelance AI and machine-learning engineer building production ML, NLP, generative AI and private/local AI systems. Berlin, Germany." url: "https://tobiassterbak.com/" --- # Tobias Sterbak — Freelance AI & Machine Learning Engineer Hi, I'm Tobias Sterbak. I design and build production AI systems — from ML and NLP applications to generative AI and private local-LLM infrastructure. My background in machine learning predates the generative-AI wave, so I use LLMs when they help rather than because they are fashionable. **Based in Berlin, Germany. Available freelance (remote or on-site).** ### At a glance - 10+ years ML & NLP - Prototype → production since 2016 - Local AI - PyCon / PyData speaker - Open-source maintainer --- ## Work I work across machine learning, language technology and software engineering — with deep technical involvement, pragmatic implementation and a preference for solutions that remain understandable, maintainable and useful in practice. ### 1. Build the AI product — Prototype → production Custom ML/AI applications, APIs and product integration — from first technical validation to a shipped feature your users actually rely on. **Stack:** PyTorch, scikit-learn, FastMCP, Unsloth (LLM finetuning), Pandas ### 2. Make it reliable — Deployment & operations Evaluation, monitoring, reproducibility and infrastructure — including open-source and self-hosted stacks that keep working after the demo. **Stack:** MLflow, LangFuse, vLLM, AWS, llama.cpp ### 3. Work with language — NLP, RAG & generative AI Search, extraction, classification, assistants and text generation — using classical NLP, transformers and LLMs where they make sense. **Stack:** Transformers, LangChain, pgvector, spaCy, LLMs, LanceDB ### 4. Build what surrounds the model — Software & data engineering Backends, databases, pipelines and supporting infrastructure — designed around maintainability and performance rather than unnecessary complexity. **Stack:** PostgreSQL, Redis, FastAPI, Docker, Linux --- ## Local AI · Private & open Run modern open models on your own infrastructure. I help select models and hardware, build local RAG systems, optimize inference and deploy maintainable self-hosted stacks. [Explore Local AI](https://tobiassterbak.com/local-ai) --- ## Projects Selected applied AI and open-source projects. Concrete systems for real workflows, alongside tools I build and maintain in the open. ### 01 — AI hiking assistant for individual route planning *(Applied AI · Personalization)* An AI assistant creates suitable hiking routes from individual preferences, taking gastronomy, personal limitations and additional requirements into account. - **Role:** Architecture · ML · Backend - **Result:** Agentic assistant built with a media-tech partner, grounded in real publisher data - **Stack:** Python · vLLM · LangFuse · LanceDB [KI.M ↗](https://medien-bayern.de/agentischer-reiseassistent-verlagsdaten/) ### 02 — Automated receipt processing for accounting *(Applied AI · Automation)* An AI-based workflow recognizes receipt contents, classifies them automatically and triggers subsequent processing steps in accounting software. - **Role:** Architecture · ML · Backend · Deployment - **Result:** Production workflow replacing previously manual data entry - **Stack:** Python · FastAPI · PostgreSQL · LLM/ML · AWS ### 03 — AI assistant for consistent campaign briefings *(Applied AI · Marketing)* The assistant helps marketing teams gather all relevant campaign information in a structured way and produces consistent briefings in the company's language and brand style. - **Role:** LLM application · Structured generation - **Result:** Consistent briefings in brand tone for a marketing team - **Stack:** LLM · Structured generation · LanceDB · Langsmith · Redis ### 04 — Legal review of rental contracts *(Client work · NLP)* NLP methods applied with dida to software that detects potential errors in rental contracts and supports legal review. - **Role:** NLP · Document analysis - **Result:** Production feature used by legal reviewers - **Stack:** Python · spaCy · Transformers [dida ↗](https://dida.do/projects/legal-review-of-rental-contracts) ### 05 — PromptMage *(Open source · LLM tooling)* A Python framework for building and managing multi-step LLM applications, with workflow management, prompt iteration, testing and evaluation. - **Role:** Maintainer · Design · Backend · Ops - **Result:** Open source · 115+ GitHub stars · MIT licensed - **Stack:** Python · FastAPI · LangChain [Visit PromptMage ↗](https://promptmage.io/) ### 06 — Biaslyze *(Open source · Responsible ML)* A toolkit for identifying bias in NLP models and creating a practical entry point for impact assessment and mitigation work. - **Role:** Maintainer · NLP - **Result:** Open source · supported by Prototype Fund - **Stack:** Python · Transformers · scikit-learn [Visit Biaslyze ↗](https://biaslyze.org/) ### 07 — OpenAndroidInstaller *(Open source · Desktop software)* A graphical installer that guides people through installing alternative Android distributions without requiring the usual command-line setup. - **Role:** Maintainer · Desktop application - **Result:** Open source · 520+ GitHub stars - **Stack:** Python · Flet [Visit project ↗](https://openandroidinstaller.org/) --- ## Experience Building machine-learning systems since 2015. A track record spanning research, competitions, production AI, open source, workshops and hands-on consulting. - **Since 2016 — Production AI & consulting:** Designing, implementing and operating ML, NLP and GenAI applications for companies — from prototype to production operation. - **Since 2020 — Open-source AI:** PromptMage, Biaslyze and other tools — supported through programs including Prototype Fund and Media Tech Lab. - **Since 2021 — Workshops & knowledge transfer:** Practical AI workshops, hackathons and technical training for teams. - **2015 — Entry into machine learning:** Research as a student assistant at Humboldt University of Berlin and applied ML through international Kaggle competitions. --- ## Talks & expertise Conference talks, tutorials and podcast conversations — PyCon/PyData speaker since 2018. - **2026 · Podcast** — Die Güte von Gen-AI-Projekten bewerten (Evaluating the quality of GenAI projects) - **2024 · Podcast** — Von Prognosen und Prompts: Data Science trifft generative KI - **2023 · PyCon DE** — How to baseline in NLP and where to go from there - **2022 · PyCon DE** — Introduction to MLOps with MLflow - **2019 · PyCon DE** — Managing the end-to-end machine learning lifecycle with MLflow - **2018 · PyData Amsterdam** — "Why Should I Trust You?" — Debugging black-box text classifiers --- ## Writing Selected evergreen technical writing from *depends on the definition* — still relevant, still referenced. - **[Causal graphs and the back-door criterion — a practical test on deconfounding](https://www.depends-on-the-definition.com/causal-graphs-and-deconfounding/)** (2023-02-19 · Causal inference) - **[How to calculate Shapley values from scratch](https://www.depends-on-the-definition.com/shapley-values-from-scratch/)** (2022-07-19 · XAI) - **[How to add new tokens to Hugging Face Transformers vocabulary](https://www.depends-on-the-definition.com/how-to-add-new-tokens-to-huggingface-transformers/)** (2022-05-12 · NLP) - **[Learning unsupervised embeddings for textual similarity with transformers](https://www.depends-on-the-definition.com/unsupervised-text-embeddings-with-transformers/)** (2021-05-24 · Transformers) [All articles ↗](https://www.depends-on-the-definition.com/) --- ## Contact Need help prototyping an AI product, putting an ML system into production or running models locally? - **Email:** [hello@tobiassterbak.com](mailto:hello@tobiassterbak.com) - **LinkedIn:** [linkedin.com/in/tobias-sterbak](https://www.linkedin.com/in/tobias-sterbak/) - **GitHub:** [github.com/tsterbak](https://github.com/tsterbak) - **Blog:** [depends-on-the-definition.com](https://www.depends-on-the-definition.com/)