Hello, I'm Jens

AI Engineer | LLM Training & Optimization | High-Performance Computing

Get in touch View projects
Jens Lücke

About Me

Who I Am

Ph.D. physicist turned AI engineer with deep expertise in High-Performance Computing. I specialize in training, fine-tuning, and optimizing large-scale AI models - from pre-training billion-parameter LLMs on distributed systems to achieving >50% performance gains via LoRA fine-tuning.

Outside of work, I compete in powerlifting - the kind of systematic, incremental progress that also defines how I approach engineering problems.

Skills & Expertise

LLM Training

Pre-training, LoRA fine-tuning, synthetic data generation

HPC & Systems

C, MPI, CUDA, distributed computing, Cerebras

AI Agents

ReAct prompting, tool use, multi-step workflows

RAG Systems

Vector search, GraphRAG, knowledge graphs

Work Experience

AI Engineer

Jan. 2024 – Present

Aleph Alpha

  • Improved VLM performance by >50% (60% → 90% accuracy) via LoRA fine-tuning and a novel synthetic data generation pipeline
  • Led pre-training of a 3B parameter German-language LLM on a Cerebras CS-3 cluster (4.4T token dataset)
  • Built GraphRAG PoC achieving 25% precision improvement over vanilla vector search on a complex legal document corpus
  • Developed ReAct-style AI agents to automate multi-step document analysis workflows with custom tool integration

Ph.D. in Theoretical Physics

Oct. 2019 – Dec. 2023

Humboldt University Berlin

  • Designed and executed large-scale QCD+QED simulations on HPC clusters across Europe, managing thousands of distributed cores
  • Authored 5,000+ line C/MPI mass-reweighting module for openQxD, improving simulation efficiency by 15%
  • Built high-performance Python pipeline (NumPy, SciPy) processing terabytes of MCMC simulation output
  • Achieved 4x GPU speedup of the Dirac operator via CUDA at an NVIDIA-sponsored hackathon
  • Member of RTG2575; taught statistical physics, quantum mechanics, and linear algebra

Full publication list on Google Scholar.

Featured Projects

convert.py 
$ python convert.py pharia-1
Loading architecture...
Mapping 142 tensors → GGUF
Quantize: Q4_K_M █████ 100%
✓ pharia-1.gguf (1.8 GB)

Pharia-to-GGUF Conversion

Reverse-engineered the Pharia model architecture for GGUF conversion. Implemented bespoke inference logic in C++ within llama.cpp for efficient quantized CPU inference.

C++ Python GGUF llama.cpp
agent.py 
$ agent "Best Wilks, 93kg?"
[Plan] Decomposing query...
[Tool] wilks_calc → 521.3
[Tool] search → openpowerlifting
[Agent] ✓ Answer ready

Agentic RAG System

ReAct-style agent for complex, multi-hop powerlifting questions that standard RAG fails on. Custom tools for structured data retrieval and web search, coordinated by a planner LLM.

Python Swarm ReAct
evolve.py 
$ evolve --pop 32 --gens 50
Gen 01 | best=0.42 ██░░░░░░
Gen 25 | best=0.87 ██████░░
Gen 50 | best=0.98 ████████
✓ Best solution: gen 47

Open Alpha Evolve

Open-source implementation of DeepMind's AlphaEvolve evolutionary algorithm, using an LLM-driven mutation operator to autonomously evolve Python code.

Python LLMs Evolutionary Algorithms

Get In Touch

Email

Run this in Python, or click to copy:

"".join(["@", "je", "ecke", ".", "ns", "ai", "lu"][i] for i in [1, 4, 0, 6, 2, 3, 5])

Location

Berlin, Germany

Connect

Hugging Face