AI model-behavior evaluation
Adversarial and edge-case testing across chat, multimodal inputs, tool use and indirect prompt injection.
Data quality · information retrieval · AI evaluation
I make information and AI systems more reliable.
I retrieve and verify information, investigate data and evidence quality, and test model behavior. When the work needs a durable system, I define the research process, test the result and keep it maintained.
Italy · EU/EEA work-authorised · sponsored international relocation · B2B engagements
Professional scope
Where I can contribute.
Information retrieval & scientific verification
Finding primary sources, tracing claims and verifying scientific evidence for public-facing work.
Data quality & provenance
Provenance, metadata, source reconciliation, validation, public-source investigation and information-quality review.
Research directories, tools & visualization
Structured records, provenance rules, maintained tools and data visualizations for recurring research-information problems.
Selected work
Selected work, shown through the actual output.
Each case shows the actual public output, my contribution and the boundary of what that work demonstrates.
01AI evaluation
Finding model failures across four evaluation surfaces.
I tested chat, image, agent and indirect prompt-injection behavior and preserved dated public evidence that keeps Proving Ground and Arena metrics separate.
- What I did
- I vary the interaction path, preserve reproduction notes and distinguish direct observations from platform labels, independent verification and model-wide conclusions.
- Result
- #74 on the Proving Ground, top 6%, with 113 platform-displayed total breaks on 29 July 2026; the Arena profile displayed rank #365, 255 submissions, 28 global unique breaks and 1,120 points.
Scope: The four visible area counters sum to 112 while the profile displays 113 total breaks. Both are reported without inferring the platform's internal aggregation, and the record is dated evaluation evidence rather than a model-wide conclusion.
Research, content & website maintenance
Public work record across evidence review, content production and publishing.
80documented published contributions55 YouTube videos · 4 articles · 21 short-form pieces
- Research
- Primary literature and scientific fact-checking
- Production
- Scripts, visuals, slides and selected packaging
- Website maintenance
- WordPress, responsive design, publishing and hosting
36.5M public YouTube views in the production environmentBrand audience; cross-platform follower totals are non-unique and time-sensitive.
View the official Entropy for Life work record ↗02Scientific fact-checking
Fact-checking and producing scientific content before publication.
Paid contractor supporting an established Italian science-communication brand through evidence review, English-to-Italian scientific localization, content production and WordPress website management.
- What I did
- Conduct recurring primary-literature research, scientific fact-checking and source-faithful English-to-Italian localization. Depending on the assignment, also develop scripts, data analyses, visualizations, slides, on-screen assets, short-form content and selected thumbnails. Designed and built entropyforlife.it in WordPress and manage responsive design, publishing, hosting, DNS, SSL and technical SEO.
- Result
- 80 documented published content contributions: 55 YouTube videos, 4 co-authored articles and 21 short-form pieces. The official work record also indexes selected thumbnail work, which overlaps with video projects and is not added to the total.
Scope: Platform metrics describe the production environment, not a personal audience. Quantified thumbnail lift is stated only when comparable analytics are available.
03Source verification & investigation
Tracing privacy, policy and evidence changes under dispute.
My public record includes consumer-genomics privacy research, corporate-source reconciliation, archival recovery, content-governance review and biomedical evidence synthesis—not a generic claim of OSINT familiarity.
- What I did
- I built and maintained the Nebula Genomics privacy record across historical sequencing relationships, changing company statements, archived privacy policies, third-party-data questions, corporate restructuring and litigation coverage. In separate cases, I recovered missing documents for a legally sensitive chronology, audited source quality and bibliometrics during a contentious moderation process, and built a syndromic-autism taxonomy from peer-reviewed literature.
- Result
- Each case is linked to an exact attributed diff, edit history or collaborative discussion record, with methods and stated limits documented on a dedicated page. Living-person names are not used as promotional headings when they add little hiring value.
Scope: The work represents public-source research and collaborative knowledge governance. It does not establish company liability, personal misconduct, editor affiliation, an independent legal judgment or original clinical research.
04Supporting research system
Building a maintained directory of research opportunities.
Yourself to Science turns scattered institutional opportunities into a public catalogue with explicit inclusion and update rules.
- What I did
- I defined the inclusion criteria, verification fields, provenance model, licensing boundaries, update process and public-data requirements, then coordinated AI-assisted implementation and maintenance.
- Result
- 55 resources indexed; as of 27 July 2026, 37 unique Wikidata items use yourselftoscience.org as a reference URL (P854). FAIRsharing, Zenodo and human- and machine-readable interfaces provide additional public records.
Scope: My technical responsibilities cover requirements, information architecture, verification, functional testing, deployment diagnosis and ongoing maintenance; implementation is AI-assisted.
Current product
Notandia works across browser and Zotero research workflows.
Notandia identifies articles from scrutinized publishers such as MDPI and Frontiers, checks Crossref/Retraction Watch for formal notices, and adds precise MDPI reference detection in Zotero. Publisher context is not an article-quality verdict.
Chrome · Edge · Firefox · Safari source · Zotero 7–9
Scientific visualization
A diagram that became a reusable public reference.
I designed this vector diagram to show overlap among monogenic conditions associated with autism, dystonia, epilepsy and schizophrenia. The Wikimedia record exposes the source file, authorship, revision history and reuse across four Wikipedia language editions.
Open the Wikimedia source record,_dystonia,_epilepsy_and_schizophrenia.svg)
More public work
Additional tools and experiments.
Working preferences
Clear goals, direct feedback and responsibility for the result.
I work best when the objective and decision rights are explicit, feedback is direct and I can follow a problem through investigation, documentation, release and maintenance.
Separate observation from interpretation
State what was directly observed before drawing a broader conclusion.
Make assumptions visible
Document the definitions, exclusions and judgments that affect the result.
Plan for maintenance
Treat updates, provenance and recovery from problems as part of the work.
Experience
Relevant experience.
The targeted CVs carry the full application detail.
Independent AI Evaluator
Independent practice · Gray Swan Proving Ground participant · Conduct self-directed adversarial testing across chat, multimodal, agentic tool-use and indirect prompt-injection settings.
Science Writer & Fact-Checker / Website Manager (WordPress)
Entropy for Life — Italy · Delivered 80 documented published content contributions for an Italian science-communication brand with 267K YouTube subscribers and 480K+ combined platform following; recurring evidence review, selected visual packaging and WordPress website management.
Founder & Project Lead
Yourself to Science™ · Founded and operate an open-source research-participation directory indexing more than 55 initiatives, with documented verification, provenance and metadata workflows.
Application documents
Start with the role you are hiring for.
Use the data-quality and research-analysis CV as the general default; the others are focused specialist documents.
Recommended general CV for analytical, data-quality and research roles
Data Quality & Research Analysis
Information retrieval, data quality, source verification, data visualization and research support.
Open CV →Specialist CV for AI evaluation and adversarial-testing rolesAI Evaluation & Model Behavior
Model testing, adversarial QA, test planning and evidence reporting.
Open CV →Integrity, trust & safety and investigation rolesTrust, Safety & Source Quality
Consumer-genomics privacy, archival OSINT, source-quality review, public-record research, content governance and investigation support.
Open CV →Editorial, community-coordination and content-quality rolesResearch, Editorial & Community Coordination
Scientific fact-checking, editorial coordination, research participation, community engagement and content production.
Open CV →Current focus
Data quality, information retrieval and AI evaluation roles.
Especially research analysis, scientific verification, model-behavior evaluation and knowledge-integrity work.