AI-Powered Skills Portrait: Revealing Student Competencies Beyond the CV

Team name: AI LEVEL

Use of AI tools :

• Chat GPT : Structuring initial ideas, rephrasing content, and conducting exploratory research on career guidance and competency frameworks.

• Claude AI (sonnet 4.6) : sources, formatting our project report.

• Gemini : comparing data, sources

• Deepl : translation (french/english)

• CapCut : pitch video, including automated voice-over and subtitling.

External feedback & contributions :

• Patricia Pierre-Gérôme - HR Engineer & Creator of PortraitScopie (Myrhmica)

• Alain Goudey - Associate Dean for Digital, NEOMA Business School

• Jérôme ONY - Expert in AI, NEOMA Business School

• AI Grand Challenge participants - Comments on bias prevention, the importance of using established occupational frameworks, and ensuring the AI encourages students to challenge their own self-assessment

• Students

Initial contribution: AI-Powered Skills Portrait: Revealing Student Competencies Beyond the CV

Final contribution: Through adaptive conversations, AI LEVEL helps students identify and express their competencies, motivations, soft skills, and hidden potential. It can rephrase, explore contradictions, detect uncertainty, and personalize questions in real time to better understand each individual. The AI then translates their answers into a Skills Profile.

When we began Phase 2, we already knew that AI Level addressed a real and widespread problem. The gap between what students are genuinely capable of and what they manage to communicate — on a CV, in an interview, in a cover letter — is not a question of talent. It is a question of language, structure, and self-awareness. We were confident in the relevance of the idea. What Phase 2 forced us to do was stress-test every dimension of it: European regulatory compliance, data governance, AI architecture, occupational framework integration, business model viability, and the ethical design of a tool that would genuinely serve students without reproducing the very biases it aimed to correct.

The list of questions we had to answer was substantial. How do we ensure compliance with European regulations? How do we handle student data responsibly and transparently? What technical architecture is both realistic and current? How do we integrate existing occupational frameworks like ROME and ESCO meaningfully, rather than decoratively? How do we build a business model that is both viable and accessible? How do we define precisely what a student would actually do with this tool — and what they would get out of it? And perhaps most critically: how do we design an AI that helps students identify real, evidenced competencies, without hallucinating skills they do not actually have, and without simply mirroring their own assumptions back to them?

These were not questions we had fully resolved at the end of Phase 1. Phase 2 was the process of resolving them.

Alain Goudey's feedback challenged us to maintain focus on the core problem we had set out to solve — and to think further ahead. His key message was direct: "Stay focused on the original problem. What is truly interesting here is that AI is changing the very nature of professions and the competencies they require. You need to design your project so that it can evolve with AI." This was a pivotal reframing. It reminded us that AI Level is not just a tool for today's labour market — it must be designed for a labour market that is itself in motion. Competencies are not static. The ROME and ESCO classifications we rely on are updated regularly precisely because the world of work changes. Our platform must be built to absorb those updates, reflect new competency demands, and help students understand not only what skills they have today, but what skills they will need to develop tomorrow. This input pushed us to articulate the project's adaptability more explicitly — both in the technical architecture (live RAG on updated ROME/ESCO data) and in the product narrative.

Jérôme Ony's feedback was the most technically transformative input we received. He identified two issues that we needed to address head-on.

The first was our technical architecture. Our initial design proposed a seven-layer pipeline — a sequential stack of discrete components: conversational frontend, NLP engine, semantic analyser, probabilistic engine, skills graph, question generator, and recommendation system. Jérôme pointed out, rightly, that this architecture reflected pre-LLM design thinking. In 2025–2026, all of these functions are natively handled by a single orchestrating LLM with a structured system prompt, augmented by Retrieval-Augmented Generation (RAG) on structured data sources. The correct architecture is far simpler and more powerful: LLM orchestrator + RAG on ROME/ESCO + structured memory + portrait generation. We revised the architecture entirely. This change is not merely cosmetic — it reflects a more honest and technically credible presentation of what AI Level actually is. The advantages over the old pipeline are real: reduced engineering complexity, fewer failure points, better semantic coherence (a single model holds the full context rather than passing outputs between modules), and a more auditable inference chain.

The second issue Jérôme raised was the risk of hallucination and over-attribution of competencies. This risk is not a minor footnote — it is the central regulatory liability of any AI-assisted skills profiling system. Inferring "leadership" from a babysitting experience, or attributing "project management expertise" from a single student event, without objective, traceable evidence, directly exposes the platform to liability under the EU Pay Transparency Directive. The Directive requires that competency evidence be documented and defensible. A Skills Portrait that is not auditable is not compliant. In response, we developed a dedicated safeguards chapter: every competency the AI identifies is assigned a confidence level (high / medium / low) based on the quantity, specificity, and consistency of supporting evidence; every competency in the final portrait is linked back to the specific conversational exchange that generated it; the student reviews and validates every element before it is finalised; and the AI explicitly frames its inferences as hypotheses throughout the conversation, inviting the student to challenge, amend, and critically interrogate every attribution. This was a structural addition to the project — not a cosmetic one. It changed how we think about the product.

We also clarified a critical architectural safeguard that had not been stated explicitly in our earlier work: all student conversations are hermetically sealed. Each session is fully isolated. No conversation data is used to fine-tune or update the underlying model. This means that incorrect, biased, or unverified student inputs cannot propagate into the model's behaviour for future users. This is essential — not only for data integrity, but for regulatory compliance and for the credibility of the Skills Portrait as an auditable document.

Exchanges with fellow participants and jury members raised three recurring concerns that shaped our final proposal. First: how do you ensure the AI does not create or reinforce bias? This pushed us to articulate the bias-prevention mechanisms in the safeguard system more precisely.

AI Level is no longer a theoretical proposal. It is a viable and realisable project. We are currently working on the web development of the platform, with integration of the PortraitScopie API from Myrhmica.

Final report: https://docs.google.com/document/d/12UlJKH8QypncKfyJM6PiTcXK3Lfl-sGv/edit

Video pitch : https://drive.google.com/file/d/1xXbwOOcQj4Az1_PZApfzQPaFeN8PWXDH/view