I build and audit the datasets frontier models are trained and tested on. Alongside that practice, I run an applied-research program that engineers frontier-grade evaluation into a 32 GB VRAM envelope – a deliberate reproducibility target rather than a hardware limit, with the scale-up tiers defined above it.
Currently AI evaluation, data quality & software engineering · Labelbox
Evidence register
Labelled-set validation – W1 agreement study, W2 contamination gap and execution-grounded correctness. Preprints v2 to follow.
Frontier-model evaluation, the data-quality discipline behind it, the tooling that keeps it repeatable – and the markets background that grounds the judgment.
Rubric-based scoring for correctness, reasoning, and instruction-following – rubrics frozen under semantic versioning and content hashes, so a verdict stays reproducible months later. Pairwise comparisons across frontier models, multi-turn prompt design for agentic coding, adversarial-robustness checks, and systematically documented failure modes.
QA and rubric-based audits of contributors' datasets for function-calling and agentic-AI projects – verifying correctness, format compliance, and consistency before delivery, and enforcing quality standards within the master-review team. Construction of training and evaluation datasets, including HFI problem sets.
Local model environments that run frontier models against real tasks – hardened sandbox, batch inference, vLLM with FP4 quantization, prefix caching, an OpenAI-compatible serving layer. Open-source tooling forked and extended as part of the deliverable: JSON support in Cerberus, multi-layer validation in Haystack.
A banking apprenticeship, a B.A. in Financial Management, and six years of self-employed quantitative trading – strategy development and backtesting across spot and derivatives markets, with full tax compliance. The discipline of positions with real money behind them, applied to evaluation.
One programme, not a portfolio: the workstreams reuse each other's components, and every claim is stated in advance as a criterion with an instrument and a threshold – then labelled measured or predicted.
The programme questionCan frontier and multi-agent AI systems be evaluated trustworthily, reproducibly, and cheaply inside a sovereign 32 GB VRAM envelope – and does the evidence still hold as that envelope scales?
Five workstreams reuse each other's components; execution grounding is the spine that runs W2 → W1 → W4, and every run inherits the same envelope.
Two worlds bridged: banking and financial management on one side, a deliberate move into data science on the other. The result is financial rigor plus a data-science toolkit – the habit of documenting a position you have to defend, applied to agentic and function-calling AI.
Labelbox · Remote · Clients: leading AI labs
MLP SE · Wiesloch, Germany · Part-time
Capstone “AI for Financial Consulting & Recruiting”: designed AI use cases to personalise financial advisory and lift qualified-applicant volume through AI targeting.
Scale AI · Freelance, Remote
BraveTrade · Self-employed, Remote
Proprietary trading across cryptocurrencies, equities, and options; strategy development and backtesting across spot and derivatives markets, plus data-driven market and risk analysis.
BraveTrade · Self-employed, Remote
Ran a commercial mining operation – hardware procurement, ROI and energy-cost optimization, uptime monitoring and tax-compliant documentation. The origin of the hardware-envelope discipline the research programme runs on.
VR-Bank eG Osnabrücker Nordland · Fürstenau, Germany
Banking operations and client work alongside the formal apprenticeship – the foundation of the finance side of this profile.
Education
Certifications
Extracurricular
A long-standing fascination with science fiction is part of what drew me to AI in the first place – and it keeps me genuinely curious about where these systems are headed. Evaluating them is how I get to look closest.
Interests
Open to roles and collaborations in AI evaluation, data quality, and ML engineering – from rubric design and dataset auditing to evaluation infrastructure for agentic systems. References available on request.