Participation au Hackatlon projet d'un Agent de Bienveillance pour la Démocratie Européenne
★
CIVIC Guardian PrivHSDAgent de bienveillance · Privacy-first
Analyse en temps réelThe Privacy-preserving Hate Speech Detection Challenge
--:--:--Concept Summary interactif · Hackathon Démocratie 2026
CIVIC GuardianDétecter le préjudice, pas la personne.
Un agent IA multilingue qui prévient, contextualise, contrecarre et transforme la haine en ligne — sans jamais simplement supprimer. Civic Guardian analyse le sens et le contexte d’un message sans avoir besoin de connaître l’identité de son auteur. Il protège la participation citoyenne sans transformer la modération en surveillance.
Notre ambition : maximiser l’utilité de la détection HSD tout en minimisant la perte de confidentialité et les risques de ré-identification.
✓ Pas de censure automatiqueLe système structure une réponse de confiance.
✓ Revue humaine obligatoirePour chaque cas à haut risque.
✓ Données strictement utilesCollectées et conservées au minimum.
✓ Preuves explicablesDes dossiers compréhensibles, jamais des verdicts opaques.
✓ MVP faisableUn démonstrateur crédible dans le temps du hackathon.
Mevlana
Cyrille
Sabrina🛡Identity Firewall
Frontière de Pareto Privacy-HSDMaximiser la qualité de détection tout en minimisant le risque de fuite.
Problem & Solution Fit
Transformer le discours haineux en démocratie active.
Le système qualifie le risque sans construire un outil de surveillance. Il prépare une réponse contextualisée, documentée et proportionnée.
Alignement avec le challenge PrivHSD
Un agent IA multilingue qui prévient, contextualise, contrecarre et transforme la haine en ligne — sans jamais simplement supprimer.
PrévenirDétecter avant l’escalade.
ContextualiserMobiliser des sources fiables.
ContrecarrerProposer une réponse compréhensible.
TransformerSuggérer une réponse civique.
Dashboard analytique
Rendre la confidentialité et l’utilité mesurables.
Les chiffres sont synthétiques : ils illustrent les indicateurs à comparer pendant le prototype et dans la note de recherche.
Évolution de la toxicité narrative
Données agrégées et anonymisées sur sept jours.
Intensité haineFausses affirmationsRéponses constructives
Scores explicables en direct
Chaque score reste interprétable et soumis à contrôle humain.
Haine ciblée91%
Harcèlement84%
Confiance89%
Fuite résiduelle0,8%
Moteur technique et innovation
Un pipeline Privacy-HSD modulaire, localement déployable et mesurable.
Le système ne demande jamais plus de données que nécessaire. Chaque étape peut être inspectée, documentée et comparée dans la note de recherche.
1. Entrée autorisée
Widget web, corpus de test ou connecteur validé. Aucun scraping incontrôlé.
2. Pare-feu d’identité
Suppression ou transformation des identifiants inutiles avant toute inférence.
3. Capsule contextuelle
Transmission du minimum conversationnel nécessaire pour interpréter le message.
4. Analyse NLP
Classification multilingue, preuves textuelles, confiance et incertitude.
5. Tests de fuite
Mesure du risque d’exposition, d’inférence et de ré-identification.
6. Contrôle humain
Recommandation lisible et décision sensible laissée à une personne autorisée.
Cœur NLP auditableModèles inspectables, règles documentées et scoring reproductible.
Orchestration auto-hébergéeWorkflow localement maîtrisé, séparé des décisions sensibles.
Stockage minimal en EuropeTraces anonymisées nécessaires, durée limitée et suppression vérifiable.
Registre de licences et SBOMInventaire des composants et conformité documentée.
Principe de confiance : pas de surveillance illimitée, pas de profilage des auteurs, pas de suivi multiplateforme des citoyens, pas de sanction automatique opaque. Civic Guardian mesure le risque avant de prétendre le réduire.
Démonstration Mevlana
Du signal détecté à une réponse proportionnée.
Le système ne bannit pas automatiquement et ne cherche pas à connaître l’auteur. Il structure un dossier de preuve explicable pour une revue humaine.
1. SignalerFil soumis dans un espace autorisé.
2. NeutraliserIdentifiants inutiles retirés localement.
3. PrioriserScores visibles, preuve et incertitude.
4. ExaminerDécision humaine pour les cas sensibles.
5. AgrégerTendances anonymisées utiles au Pacte.
Publication citoyenne · Mevlana« Chacun devrait pouvoir prendre part au débat public sans craindre les insultes ni le harcèlement. »
Faux compte · message discriminatoire masqué« Retourne chez toi… [contenu haineux masqué] … vous profitez tous du système. »
Réponse civique suggérée« Ce message cible une personne et un groupe sur des critères discriminatoires. Revenons aux faits et au sujet du débat. Une revue humaine peut maintenant examiner ce dossier. »
Carte thermique · signaux par heure et canal autorisé
Les tendances agrégées aident à détecter les fenêtres d’escalade prioritaires sans exposer les individus.
Répartition des signaux
Contenus nécessitant une revue.
Harcèlement 38%Discrimination 27%Désinformation 21%Escalade 14%
Impact · New Democratic Pact
Protéger les personnes sans affaiblir le débat démocratique.
Civic Guardian transforme la modération en assistance démocratique : il explique, contextualise, désamorce et documente, sans automatiser la censure.
Pour les citoyens
Participer plus sereinement au débat public, reconnaître la manipulation et recevoir une contextualisation proportionnée.
Pour les modérateurs et ONG
Prioriser les cas urgents grâce à des signaux explicables, des preuves traçables et une revue humaine structurée.
Pour les institutions et chercheurs
Observer des tendances agrégées et anonymisées pour éclairer les politiques publiques sans exposer les individus.
Minimum Viable Product
Démontrer l’efficacité du modèle sans bannir ni identifier l’auteur.
Le MVP est ambitieux, mais livrable dans le temps du Hackathon.
Conversation live
Importer un fil de démonstration.
Anonymisation
Neutraliser localement les identifiants.
Scores explicables
Afficher haine, escalade, confiance et fuite.
Dossier de preuve
Afficher signaux utiles et sources.
Revue humaine
Valider toute décision sensible.
Équipe · Avantage Digital
Trois expertises complémentaires.
Le droit fixe la limite. La technologie la mesure. Le design la rend visible. La démocratie gagne en confiance.
Mevlana Shaiti
Consultante en droit politique · DéfenseureTraduit les droits fondamentaux en règles opérationnelles et recours humains.
Cyrille Bersot
Fondateur d’Avantage Digital · Technologue IAConçoit le pipeline Privacy-HSD, l’orchestration et le prototype.
Sabrina Hohengarten
Fondatrice de Rebelles France · UX & VisualisationRend les garanties visibles, compréhensibles et actionnables.
Alignement politique
Protéger le débat public sans transformer la modération en surveillance ou profilage.
Innovation technique
Pare-feu d’identité, analyse contextuelle et dossier de preuve explicable.
Synergie multidisciplinaire
Droit, technologie et design coopèrent pour rendre les droits concrets.
⭐ Proposition de valeur finale : Civic Guardian transforme le discours haineux en démocratie active : il détecte, explique, contextualise, désamorce et structure une réponse de confiance. Il protège les personnes sans créer un nouvel outil de surveillance.
Hackatlon 2026 Civic Guardian A.I agent Project for Europeen Democracy
Civic Guardian PrivHSD
Civic AI support agent · Privacy-first
Real-time analysis
Confidential
--:--:--
Privacy-HSD
Human Rights by Design
Auditable prototype
New Democratic Pact
Detect harm, not the person.
Civic Guardian PrivHSD is a multilingual hate-speech detection system designed to analyse the meaning and context of a message without needing to know the author’s identity. It protects civic participation without turning moderation into a surveillance tool.
✓ Identity not required
✓ Minimal context
✓ Explainable scores
✓ Human validation
⏱ PrivHSD application · Democracy Hackathon · Strasbourg · 17-19 June 2026
🛡Identity Firewall
Identifiers neutralisedName, profile, geolocation and metadata removed.
Context CapsuleOnly the strictly necessary context is transmitted.
Identity-agnostic analysisThe content is assessed without profiling its author.
Privacy-HSD Pareto frontier
Maximise detection quality while minimising leakage risk
A different approach
Fight hate without building a surveillance tool.
Privacy protection is not added at the end of the project. It determines what enters the engine, what is excluded and what remains subject to human review.
01
Minimise
Remove unnecessary identifiers before inference and strictly limit retention.
02
Understand
Analyse the meaning of the message and the minimum conversational context that is genuinely necessary.
03
Measure
Compare HSD performance with residual leakage, inference and re-identification risks.
04
Explain
Display the main reasons, uncertainty and safeguards applied before any sensitive decision.
Fit with the PrivHSD challenge
Address the core problem: detect hate without exposing individuals.
The project is designed to maximise the Privacy-HSD trade-off: maintain useful detection while reducing profiling, re-identification and malicious doxing risks.
01 · Problem-solution fit
Analyse statements, not profiles.
The engine receives sanitised content and a limited Context Capsule. The author’s identity, precise location and cross-platform history are not required.
02 · Integrated innovation
AI/NLP, law and UX operate as a single system.
The Identity Firewall, multilingual semantic classifier, adversarial privacy tests, explainable scores and human validation are integrated from the architecture stage.
03 · Impact potential
A demonstrable prototype, measurable research and a startup pathway.
The proof of concept can be tested on synthetic or governed datasets, then deployed with NGOs, moderated communities and institutional stakeholders.
04 · Human Rights by Design
Legal safeguards become visible features.
Data minimisation, short retention, visible uncertainty, human review, freedom of expression and the audit trail are not merely legal notices.
A genuinely multidisciplinary response.
Compliance does not come after the code: Mevlana defines the legal boundaries, Cyrille builds the auditable engine and Sabrina turns safeguards into understandable interactions. Their work converges towards a useful, proportionate and demonstrable tool.
Commitment to auditability. The prototype clearly separates the scientific core from the rest of the stack: controllable components, licence register, software inventory, reproducible scoring and authorised connectors. Proprietary APIs may serve as entry points, but the central logic must remain inspectable and responsible.
Multidisciplinary synergy · 20% of the assessment
Three areas of expertise. One responsible architecture.
The PrivHSD challenge requires more than a high-performing engine. Law, engineering and user experience must work together so that privacy protection remains measurable, verifiable and understandable.
⚖ Policy & Law
Guardian of purpose limitation
Mevlana Shaiti
Advocate · Policy & Law
She translates Council of Europe principles into operational rules: authorised purposes, excluded data, human review and safeguards against surveillance creep.
- Convention 108+, GDPR and ECHR
- Freedom of expression and proportionality
- Digital responsibility and legal audit
🧠 Engineering & AI
Engine designer
Cyrille Bersot
Technologist · Engineering & AI
He designs the Identity Firewall, the Context Capsule, the auditable NLP pipeline and the indicators that genuinely measure the trade-off between utility and privacy.
- Self-hosted and modular architecture
- Reproducible scoring and adversarial testing
- Licence register and auditable components
✦ UX & Visualisation
Translator of complexity
Sabrina Hohengarten
Designer · UX & Visualisation
She makes human rights visible in the interface: anonymisation status, score explanations, uncertainty, audit trail and human validation.
- Accessible UX for non-specialists
- Data visualisations that support decision-making
- Understandable Human Rights by Design
Collective working method
Law sets the limit. Technology makes it measurable. Design makes it visible.
This continuous dialogue prevents two frequent failures: technically brilliant but legally dangerous innovation, or abstract compliance that is impossible to understand in real-world use.
01⚖️Policy & LawDefine the purposes, boundaries and safeguards.
02🧠Engineering & AICode, measure and document safeguards.
03✦UX & VisualisationMake rights understandable and actionable.
04🏛️Democratic impactProtect people without weakening debate.
Auditable technical engine
A modular, locally deployable and measurable Privacy-HSD pipeline.
The system never requests more data than necessary. Every step can be inspected, documented and compared in the research note.
1. Authorised input
Web widget, test corpus or validated connector. No uncontrolled scraping.
2. Identity Firewall
Removal or transformation of unnecessary identifiers before any inference.
3. Context Capsule
Transmission of the minimum conversational context required to interpret the message.
4. NLP analysis
Multilingual semantic classification, textual evidence, confidence and uncertainty.
5. Leakage tests
Measurement of exposure, attribute-inference and re-identification risks.
6. Human review
Readable recommendation, audit trail and sensitive decision left to an authorised person.
Trust principle: no unlimited surveillance, no author profiling, no cross-platform tracking of citizens, no opaque automated sanctions. Civic Guardian measures risk before claiming to reduce it.
Conversational demonstration
The Identity Firewall acts before the AI reads the message.
The sequence starts when this section reaches the centre of the screen. The scenario illustrates anonymised example messages and shows what the engine retains - and above all what it refuses to use.
f Facebook
Example of an analysed conversation
MevlanaPublic post · 12 min ago · 🌐
“Thank you to the residents who took part in our civic workshop this weekend. The discussions were simple, warm and useful. This is how we build a city where everyone can find their place.”
👍 Like💬 Comment↗ Share
Vérité_Réveillée_25 · suspicious profile“Always the same people taking advantage of the system. Roma automatically receive benefits and then come to lecture us. Go back to where you came from.”
Mevlana“I would like us to speak respectfully. Could you provide an official source for this claim? My background does not justify this type of comment.”
Civic AI support agent suggestion · pending validation“Social benefits are not automatically awarded according to a person’s background. They depend in particular on residence, circumstances and resources. This comment also contains an exclusionary injunction directed at a person. A factual and respectful response helps restore a constructive framework for discussion.”
Illustrative portrait. The messages reproduce anonymised example situations. The system assesses content and its context; it does not claim to determine an author’s real identity.
Privacy-HSD metrics
The privacy promise becomes a measurable result.
The displayed data are synthetic. They illustrate the research dashboard used to compare engine utility before and after data sanitisation.
Macro‑F1 HSD
0,00
↑ Useful detection after minimisation
Identifiers removed
0%
✓ Before inference
Residual leakage
0,0%
↓ Limited exposure
Blocked re-identification attempts
0%
↑ Adversarial testing
Pareto frontier · privacy vs utility
Seek the best balance rather than optimise an isolated score
Privacy-HSD frontierTarget balance point
Privacy tests
Metrics planned for the research note
HSD precision89%
HSD recall85%
PII exposure prevented99%
Re-identification blocked98%
Verifiable deletion100%
Signal intensity · by hour and authorised channel
The heatmap aggregates anonymised trends and helps identify priority escalation windows
Content requiring review
Aggregated distribution of processed signals
Harassment · 38%Discrimination · 27%Disinformation · 21%Escalation · 14%
Human Rights by Design · New Democratic Pact
Human rights are not a footnote. They become features.
Civic Guardian is designed to protect dignity, privacy and freedom of expression simultaneously. Each legal principle is translated into a visible, understandable and auditable control.
Convention 108+
Modernised data protection
Minimisation, limited purposes, proportionate processing and control mechanisms adapted to digital technologies.
GDPR
Strictly necessary data
Unnecessary identifiers are neutralised before inference. Remaining traces are time-limited and deletable.
ECHR · Articles 8 and 10
Privacy and freedom of expression
The system distinguishes hate speech from lawful controversy, satire and legitimate democratic disagreement.
CETS No. 225
AI, democracy and the rule of law
Transparency, accountability, a risk-based approach and human oversight structure the system lifecycle.
Identity FirewallNames, profiles, precise locations and unnecessary metadata are removed or transformed before analysis.Concrete safeguard: privacy from the point of entry.
Minimal Context CapsuleThe engine receives only the context essential to interpret a message, never a complete file on the author.Concrete safeguard: minimisation and proportionality.
Visible confidence and uncertaintyA score never becomes an automatic truth. The interface shows the limits of the analysis and the signals actually used.Concrete safeguard: transparency and contestability.
Proportionate human validationThe engine suggests, explains and documents. An authorised person remains responsible for sensitive decisions.Concrete safeguard: human responsibility.
Audit trail and verifiable deletionThe applied safeguards, retention period and deletion of traces can be verified.Concrete safeguard: institutional trust.
Respect the spirit of the New Democratic Pact
Strengthen democracy without building surveillance infrastructure.
Democratic participation cannot be safe if citizens must choose between enduring hate or being profiled in order to be protected. Civic Guardian rejects this false dilemma.
✓ Safe participation
✓ Protected dignity
✓ Preserved freedom of expression
✓ Contestable decisions
✓ Institutional transparency
✓ Responsible technology
What the interface shows the user
Make rights visible in under ten seconds
“Identifiers removed before analysis”
The status confirms that the Identity Firewall worked before inference.
Residual leakage risk
The remaining exposure level and the applied tests are displayed.
Textual evidence and uncertainty
The user understands why an alert was produced and can discuss it.
Human-validation status
The person knows whether an action has been proposed, reviewed or validated.
What Civic Guardian refuses to become
Stop purpose creep before it becomes routine
Author profilingProhibited
Inference of political affiliationProhibited
Cross-platform tracking of citizensProhibited
Ranking or scoring of peopleProhibited
Opaque automated sanctionProhibited
Digital responsibility: Civic Guardian assesses a message and its strictly limited context. It does not seek to discover who a person is. It documents the applied safeguards, makes uncertainty visible and maintains human oversight for every sensitive decision.
Impact · Translate the New Democratic Pact into a concrete tool
Help citizens. Protect democratic debate. Restore trust.
Civic Guardian transforms raw signals into proportionate, privacy-respecting information. The objective is not to control debate, but to allow everyone to participate without enduring hate or giving up their rights.
Citizens
Understand hate signals, receive a clear explanation and regain a safer discussion space without being profiled.
Protected democratic participationModerators, NGOs and associations
Prioritise alerts, contextualise situations and intervene consistently through explainable scores.
Proportionate and documented actionDemocratic institutions
Receive aggregated and structured trends to inform public consultations without unnecessarily exposing people.
Transparency and institutional trustResearch and public policy
Compare HSD performance with residual leakage risks and share reproducible results to improve practices.
Evidence-based decisionsFrom an individual signal to democratic improvement
Create actionable information without turning people into exploitable data.
01Signal detectedA message or conversation requires analysis.
→
02Identity neutralisedUnnecessary personal data are removed before inference.
→
03Qualified contextThe engine provides an explainable score and an uncertainty level.
→
04Proportionate responseA human reviews, contextualises and selects the appropriate action.
→
05Aggregated insightsAnonymised trends inform policies and consultations.
Realistic roadmap
Start small, demonstrate scientifically, deploy with care.
1
Strasbourg prototype
Example conversation, Identity Firewall, Context Capsule, NLP classification, leakage tests, utility/privacy comparison and explainable dashboard.
2
Governed pilot
Tests on synthetic or governed datasets, followed by experimentation with NGOs or authorised moderated communities under human oversight.
3
Civic-tech startup and research
Modular moderation-support service, privacy-respecting aggregated observatory and contribution to a scientific publication.
“Protecting democracy does not mean choosing between security and freedom. It means designing technology responsible enough to defend both.”
Confidential · Protected mock-up and concept · Property of Avantage Digital and Cyrille Bersot
Dossier de candidature déposé ( English)
★
CIVIC Guardian PrivHSDDemocratic Benevolence Agent · Concept Summary
The Privacy-preserving Hate Speech Detection ChallengeCIVIC GuardianDetect harm, not the person.
A multilingual tool that detects hate speech without depending on the author's identity. Civic Guardian analyses harm, contextualises the situation and structures a civic response without turning moderation into surveillance.
Our ambition: maximise HSD utility while minimising privacy loss and re-identification risks.
SabrinaCyrilleMevlana
Identity Firewall
Identifiers removedName, profile, links and metadata are excluded.
Minimal contextOnly what is strictly needed reaches the engine.
Explainable outputScore, evidence, uncertainty and human review.
Privacy-HSD Pareto frontier
Maximise detection quality while minimising leakage risk.
🌍
Scope46 CoE member states
🔒
DataMinimisation by default
🤖
AI engineNLP · RAG · semantic LLM
👥
UsersCitizens · NGOs · Institutions
01
Problem & Solution Fit
Beyond deletion: act on social dynamics while protecting people.
The dual democratic risk
Online hate combines harassment, ideological framing, false claims and emotional contagion. It erodes social cohesion. Yet an overly intrusive tool can also become an instrument of profiling, tracking or re-identification.
×Do not analyse more data than necessary.
×Do not track people across platforms.
×Do not confuse moderation with surveillance.
Our response: analyse the message, not the person
The system neutralises identifiers before inference, keeps a limited context capsule and produces an explainable evidence bundle. Civic Guardian does not automate censorship; it structures trusted response.
✓Human review remains mandatory for high-risk cases.
✓Only necessary data is collected and retained.
✓Evidence bundles, not black-box verdicts.
⭐ Challenge alignment
Turn hate speech into active democracyA multilingual AI agent that prevents, contextualises, counters and transforms online hate - without ever simply deleting it.
PreventDetect before escalation.
ContextualiseUse reliable sources.
CounterProvide an understandable rebuttal.
TransformSuggest a civic response.
02
Technical engine and innovation
A modular, locally deployable and measurable Privacy-HSD pipeline.
Inspectable innovation at every step
The system never requests more data than necessary. Every step can be inspected, documented and compared in the research note.
1. Authorised input
Web widget, test corpus or validated connector. No uncontrolled scraping.
2. Identity Firewall
Removal or transformation of unnecessary identifiers before any inference.
3. Context capsule
Transmit the minimum conversational context needed to interpret the message.
4. NLP analysis
Multilingual semantic classification, textual evidence, confidence and uncertainty.
5. Leakage tests
Measure exposure, attribute-inference and re-identification risks.
6. Human oversight
Readable recommendation, audit log and sensitive decision left to an authorised person.
🧠 Auditable NLP core
Inspectable multilingual models, documented rules and reproducible scoring.
⚙️ Self-hosted orchestration
Locally controlled workflow, separated from the scientific engine and sensitive decisions.
🇪🇺 Minimal storage in Europe
Strictly necessary anonymised traces, limited retention and verifiable deletion.
📋 Licence register and SBOM
Component inventory, applicable licences and documented compliance.
Trust principle: no unlimited surveillance, no author profiling, no cross-platform citizen tracking and no opaque automated sanction. Civic Guardian measures risk before claiming to reduce it.
💬 Demonstration: Mevlana case
Citizen post: “Everyone should be able to take part in public debate without fear of insults.”
Fake account: “Go back home... [discriminatory message masked].”
Suggested civic response: “Your message targets a person and a group on discriminatory grounds. Let us return to the facts and the subject of the debate.”
📊 Explainable evidence bundle
Identifiers removed before analysis. No author profile is created.
Targeted hate
high
Escalation risk
review
Residual leakage
measure
⚖️ Human Rights by Design - 5 visible safeguards
1. PrivacyIdentity Firewall before inference.
2. MinimisationStrictly necessary context.
3. ExplainabilityVisible scores, evidence and sources.
4. UncertaintyConfidence bands and hesitation labels.
5. Human recourseNo automatic sensitive action.
Macro-F1HSD performance to compare.
Δ utilityBefore / after anonymisation.
Residual PIIExposure after processing.
Re-identificationAttack success rate.
5. Technical stack: separate the scientific core from auxiliary toolsLocal, auditable and privacy-first by design.
Open and measurable architecture
InterfaceAuthorised connectors, dashboard and human validation.
OrchestrationSelf-hosted n8n automates already sanitised workflows on a controlled server.
Auditable corePython, Presidio / spaCy, Transformers and FastAPI: local processing, inspectable NLP and reproducible benchmark.
DataEncrypted PostgreSQL, limited retention, verifiable deletion and licence register.
Governance rules
✓ The HSD core remains local, separable and auditable.
✓ Airtable is reserved for project management or synthetic data.
✓ A proprietary LLM may intervene only after anonymisation, for reformulation - never to decide the HSD score.
✓ A software inventory and licence register document every component.
ReceiveMinimal payload
SanitiseIdentifiers removed
ClassifyUseful context capsule
AggregateAnonymised metrics
DeleteVerifiable TTL
03
Impact
Turn the principles of the New Democratic Pact into a practical tool for real-world use.
For citizens
Participate more safely in public debate, recognise manipulation and receive proportionate contextualisation.
For moderators and NGOs
Prioritise urgent cases through explainable signals, traceable evidence and structured human review.
For institutions and researchers
Observe aggregated and anonymised trends to inform public policy without exposing individuals.
Democratic public value
Civic Guardian turns moderation into democratic assistance. It does not merely remove harmful content: it explains, contextualises, de-escalates and educates. Privacy protection becomes a condition for civic participation.
Intersectional lens
The dashboard can isolate forms of targeting: gender, perceived origin, religion or orientation. Views are aggregated to identify coordinated campaigns without creating individual profiles.
04
Minimum Viable Product
Demonstrate the effectiveness of the analytical hate-speech detection model without needing to ban or identify the author.
The proof to deliver during the Hackathon
From a demonstration thread, Civic Guardian must recognise hate speech, neutralise unnecessary identifiers, display understandable signals and suggest a proportionate response subject to human review.
05
Team - Avantage Digital
Three complementary profiles: rules, engine and user experience.
Mevlana Shaiti
⚖️ Political law consultant · AdvocateTranslates fundamental rights into ethical rules, usage limits and human-recourse requirements.
Cyrille Bersot
🤖 Founder of Avantage Digital · AI TechnologistDesigns the Privacy-HSD pipeline, n8n orchestration, prototype and API integrations.
Sabrina Hohengarten
🎨 Founder of Rebelles France · UX & VisualisationMakes safeguards visible, understandable and actionable for users and the jury.
Law sets the limit
Technology measures it
Design makes it visible
Democracy gains trust
✓
Fit with the principles of the New Democratic Pact
Policy alignment | Technical innovation | Synergy
Policy alignment
Protect public debate without turning moderation into surveillance or profiling.
Technical innovation
Identity Firewall, contextual analysis and an explainable evidence bundle to act without identifying the author.
Multidisciplinary synergy
Law, technology and design work together to make rights concrete, visible and applicable.
What Civic Guardian refuses to become
No mass surveillance, no hidden cross-platform tracking, no re-identification tool, no citizen scoring and no opaque automated sanction.
⭐ Final value proposition
Civic Guardian turns hate speech into active democracy: it detects, explains, contextualises, de-escalates and structures trusted response. It protects people without creating a new surveillance tool.
✅ Five commitments readable in one second
No automated censorshipThe system structures trusted response.
Mandatory human reviewFor every high-risk case.
Strictly necessary dataCollected and retained to a minimum.
Explainable evidenceBundles, never black-box verdicts.
Feasible MVPCredible demonstrator in 2.5 days.
★
Turn hate speech into active democracy
Civic Guardian helps detect, explain, contextualise and de-escalate online attacks. It protects people without requiring the author's identity and without automating censorship.
1 / 6
