Building AI agents for complex reality.
I design, build, and deploy production-grade AI solutions — from RAG systems at scale to autonomous agents that solve hard business problems.
Core Expertise
The technology stack I leverage to solve problems.
Advanced RAG
Beyond basic retrieval. I build multi-hop systems, hybrid search, and graph-based retrieval using LangChain and Vector DBs.
Agentic Workflows
Building autonomous agents that can plan, reason, use SQL tools, and execute Python code to answer complex queries.
Fine-Tuning
Adapting open weights (Llama 3, Mistral) to specific domain language and constraints for maximum performance.
Prompt Engineering
Systematic prompt optimization and evaluation (Evals) to ensure reliability and reduce hallucinations.
Frontier Research Focus
- Spatial Intelligence Moving beyond text to 3D understanding using Gaussian Splatting and NeRFs for scene reconstruction.
- World Models & Video Exploring JEPA architectures and generative video models to teach agents physics and causality.
- Embodied Agents Sim-to-Real transfer learning for autonomous navigation and manipulation in unstructured environments.
Selected Work
Real results from deployed systems.
ROSALIND
The R&D Knowledge Engine
Retrieval-Augmented Generation over PubMed articles to support literature review and hypothesis exploration.
AURELIA
Autonomous Orchestration
A coordinated group of agents: SQL agent, RAG agent, and Python agent working together to generate reports.
RESPONSIBILIBOT
The Guardrail System
Tailored AI assistants with strict governance, PII masking, and hallucination checks.
Philosophy
How I approach building AI products.
Frame the Problem
AI isn't magic. I start by understanding the data constraints and defining what "good" output looks like.
Prototype Rapidly
I build functional prototypes early to validate feasibility using tools like Databricks or local LLMs.
Validate & Evaluate
Evaluation is key. I implement automated evals to measure accuracy, latency, and faithfulness.
Productionize
Moving from notebook to production means handling errors, rate limits, and monitoring drift.
Games
Playful neural experiments built on expressive backgrounds.
Zip Packing
Bundle embeddings into perfect sequences.
Neural Constellations
Steer the ambient particles and grow a living cluster.
Gradient Ascension
Ascend the layers. Optimization as a vertical sprint.
Gradient Descent
Ride the loss landscape and dodge saddle traps.
Diffusion Playground
Watch the sampler run inline without leaving the page.
The Discriminator
Spot adversarial data as it evolves across epochs.
The Lab
Interactive simulations designed to explain AI concepts to non-technical stakeholders.
About Me
Driven by curiosity and code.
The Background
As a Forward Deployed Engineer, I operate at the edge where research meets reality. I stay deeply connected to the latest developments in model architecture and safety, effectively translating frontier capabilities into reliable systems. My focus extends beyond code to the broader impacts of AI, ensuring we build solutions that are not just technically sound, but ethical and sustainable.
Erik Röntgen
Forward Deployed Engineer | AI Engineer Based in GermanyContact
Interested in collaborating or discussing AI strategy?
Let's build something.
I am always open to discussing new opportunities, technical challenges, or just geeking out over the latest paper.