Adversarial AI Evaluation of Frontier Language Models
Adversarial AI evaluation stress-tests models with hard, purpose-built tasks to find where they fail. As an independent contractor for Mercor, I evaluate frontier language models on scientific reasoning in bioinformatics: I design experimental tasks and synthetic clinical datasets, analyse failure modes and build reproducible multi-agent harnesses and execution environments.
Book a discovery callWho is it for?
- AI teams that need to know where their models break in scientific domains
- Companies deploying LLMs in healthcare workflows
- Groups that need reproducible evaluation harnesses
What you get
- Evaluation task design with clear success criteria
- Synthetic clinical datasets for testing
- Failure-mode analysis and mitigation strategies
- Reproducible harness and documentation
How it works
- 1Define the capability to test
- 2Design adversarial tasks and data
- 3Run the evaluation and classify failures
- 4Report limits and mitigation options
Typical scope and pricing
Every engagement starts with a 30-minute discovery call. Pricing is custom and depends on scope; the Discovery, Build and Scale packages describe the usual ways of working. I work from Valencia, remotely with teams across Europe and the USA, in English and Spanish.
Related case studies
Frequently asked questions
Who do you evaluate models for?
I work as an independent contractor for Mercor. I do not share benchmark details beyond this public description.
Which domains?
Scientific reasoning in bioinformatics: genomics, transcriptomics and epigenomics.
Can this help my own LLM project?
Yes. The same approach (task design, synthetic data, failure-mode analysis) applies to LLM features in healthcare workflows.