Lausanne, Switzerland · building QbitDojo

Medhane Hadush AI security, engineered.

$role --current

Cybersecurity engineer with hands-on depth in AI/LLM security: adversarial attack-pattern research, judge-model evaluation pipelines and training infrastructure mapped to the OWASP LLM Top 10. Five-plus years across a national security agency, a FINMA-regulated Swiss bank and independent agentic-AI engineering. Off the clock I build voice-AI and computer-vision side projects as RioticGamer.

Launch terminal Download CV Let's talk

OWASP LLM Top 10 Multi-agent systems Splunk · MITRE ATT&CK FINMA · GDPR Red-teaming

Keynote · Swiss Cyber AI Conf 2026

385 AI security labs shipped

29.9k+ views · PedTalk AI

years in security

AI attack/defense labs

views on PedTalk AI

0/10

OWASP LLM Top 10 covered

01 terminal

Talk to the portfolio

A real shell, not a screenshot. Type a command or click a chip; every answer is pulled from my CV and project data.

medhane@hadush: ~ — zsh

● online--:--

02 about

Security engineer who builds

I work at the intersection of offensive AI research and defensive engineering. Over the last two years I designed and shipped a production LLM security training platform mapped to the OWASP LLM Top 10, with 385 interactive attack/defense labs, automated adversarial grading and a cascaded judge-model evaluation pipeline that enforces structured output, degrades gracefully and respects token budgets.

Before that I operated inside environments where mistakes are expensive: a national security agency (INSA, Ethiopia) and a FINMA-regulated bank (BCV, Lausanne), tuning Splunk detections against MITRE ATT&CK, improving endpoint intrusion detection and leading fraud investigations.

I'm comfortable in high-discretion, compliance-heavy settings, and I enjoy translating deep technical security concepts for non-technical audiences, most recently as a keynote speaker at the Swiss Cyber AI Conference 2026.

Agency → Bank → IndependentINSA national security · BCV FINMA-regulated banking · QbitDojo

Guest instructor, Powercoders BernLLM security & red-teaming sessions for the ICT bootcamp I graduated from

English · French · Amharic · TigrinyaBased in Lausanne, working across Swiss and international teams

now

Shipping new labs and judge-model improvements for QbitDojo
Running LLM red-team sessions at Powercoders (Bern)
Building FleetHawk: silent-failure detection for AI agent fleets
Maintaining PedTalk AI, my real-time voice-AI side project
Pushing French from B1 upward

PythonTypeScriptBashSQLLinux 5+ yrsDockerPostgreSQL

03 skills

A 3D map of what I work with

Drag to rotate, hover a node, filter by domain. The core is the AI-security work; everything else orbits it.

skill-graph · 42 nodes drag · hover · filter

AI / LLM security

Adversarial prompt engineering & jailbreak researchexpert

Prompt-injection attack & defense patternsexpert

Judge-model evaluation architecturesadvanced

OWASP LLM Top 10 threat modellingexpert

Agentic AI & engineering

Multi-agent orchestration & fleet operationsadvanced

Tool-use guardrails, structured output, token budgetsadvanced

Python · TypeScript · Next.js · PostgreSQL · Dockeradvanced

Gemini · Claude · OpenAI APIs, RAG, evalsadvanced

Detection & incident response

Splunk SIEM · detection-as-code · use-case engineeringadvanced

MITRE ATT&CK mapping · log analysisadvanced

Triage, containment, forensics, fraud investigation (RIO)advanced

HIPS · IDS/IPS · firewalling · Nginx hardeningproficient

Governance, identity & infrastructure

FINMA · GDPR · NIST CSF · OWASP Top 10advanced

Multi-role RBAC · auth flows · session securityadvanced

EU-hosted cloud · credential rotation · bot protectionadvanced

Linux (5+ yrs) · Bash · Git · Wireshark fundamentalsadvanced

04 ai security & agentic ai

Attack patterns in, defenses out

The core of my current work: turning adversarial research into training, evaluation pipelines and agent systems that hold up under pressure.

OWASP LLM Top 10 · 2025

Threat coverage across every category

385 labs · automated grading

LLM01Prompt InjectionDirect & indirect injection labs; the judge pipeline's core test case.

LLM02Sensitive Information DisclosurePII and secret leakage scenarios with output scanning.

LLM03Supply ChainPoisoned models, plugins and dependencies in the QAISP exam bank.

LLM04Data & Model PoisoningTraining-data tampering and backdoor behaviour labs.

LLM05Improper Output HandlingXSS / SQLi through model output; structured-output enforcement.

LLM06Excessive AgencyLeast-agency design and tool permissioning for agents.

LLM07System Prompt LeakageExtraction attacks and hardening patterns.

LLM08Vector & Embedding WeaknessesRAG poisoning and cross-tenant retrieval scenarios.

LLM09MisinformationHallucination handling and grounding checks.

LLM10Unbounded ConsumptionToken-budget safeguards and denial-of-wallet defenses.

Judge-model evaluation pipeline

Cascaded scoring, live

Multi-model judges score each defense across three weighted dimensions. Structured output is enforced, degradation is graceful and every call has a token budget. Watch a few evaluations run.

input ›…

stage 1

Guardrail

pattern & policy

stage 2

Judge A

fast triage

stage 3

Judge B

deep reasoning

stage 4

Judge C

tie-break

output

Verdict

…

blocks attack

–

preserves function

–

minimality

–

Agentic AI

Fleets of agents, kept honest

I've been building and operating multi-agent systems for over a year: orchestration, tool permissioning, evals and the monitoring that catches agents failing silently.

Orchestratorplans · routes · verifies

Plannerscopes tasks

Coderimplements

Reviewerdiffs & lint

Testerruns suites

Securitythreat checks

Docswrites changelog

Multi-agent orchestrationRan a 7-agent autonomous development fleet (planning, coding, review, testing, security, docs) for day-to-day product work.
FleetHawkOpen-source CLI that detects silent failures in AI fleets: wrong models, stale configs, ghost monitors, agents that stopped producing output.
PedTalk AILLM-driven NPC conversations with context, memory and spatial voice inside a AAA game; close to 30k views on the public release.
Guardrails by designLeast agency, deterministic paths before model calls, raw telemetry and logs never sent to a model unfiltered, human approval on anything irreversible.

FleetHawk on GitHub

05 experience

Where I've operated

Independent technical work · AI security engineering

Jan 2025 – present

QbitDojo · Powercoders · Swiss Cyber AI Conference

Lausanne, Switzerland

Designed and shipped a production LLM security training platform mapped to the OWASP LLM Top 10: 385 interactive attack/defense labs with automated adversarial grading.
Architected a cascaded judge-model evaluation pipeline on EU-hosted cloud infrastructure with structured-output enforcement, graceful degradation and token-budget safeguards.
Implemented multi-role RBAC, GDPR-compliant consent, disposable-email blocking, bot protection and credential-rotation hygiene.
Keynote speaker, Swiss Cyber AI Conference 2026 (Lugano): "Hacking LLM Applications: Attack Patterns, Defense Strategies, and Why AI Security Training Matters."
Partner and guest instructor at the Powercoders ICT Bootcamp (Bern), delivering LLM security and red-teaming sessions.

LLM securityAgentic AINext.js · PostgreSQLGDPR · RBAC

Technical Security Coordinator

Aug 2024 – Dec 2024

Banque Cantonale Vaudoise (BCV)

Lausanne, Switzerland · FINMA-regulated banking

Operated within high-availability, strict-compliance banking infrastructure with zero-tolerance incident response.
Enhanced the HIPS methodology, improving endpoint intrusion-detection accuracy across the bank's estate.
Developed the RIO incident-response methodology, adopted by the security team.
Led fraud incident investigations, reducing financial exposure and improving response SLAs.

HIPSIncident responseFINMAFraud investigation

Cybersecurity Engineer (internship)

Aug 2023 – Aug 2024

Banque Cantonale Vaudoise (BCV)

Lausanne, Switzerland

Tuned SIEM detection use cases in Splunk and integrated MITRE ATT&CK into detection, response and mitigation workflows.
Proposed an enhanced incident-response methodology that the team adopted.
Conducted fraud investigations across banking customer incidents.

SplunkMITRE ATT&CKDetection engineering

Cybersecurity Analyst

Jan 2018 – Sep 2020

Information Network Security Agency (INSA)

Addis Ababa, Ethiopia · national security agency

Applied machine-learning techniques to automate cryptanalysis tasks: early applied-ML work in a national-security context.
Built a web-based internal security training platform with cloud storage integration, used by agency staff.
Developed and delivered national cybersecurity awareness training programmes.

Applied MLCryptanalysisSecurity awareness

06 projects

Things I've shipped

AI security training platformlive

QbitDojo

Hands-on AI security training mapped to the OWASP LLM Top 10: prompt injection, jailbreaking, data poisoning, output handling and more, each lab graded automatically by adversarial judge models. Built with privacy by design on EU-hosted infrastructure.

385interactive labs

10/10OWASP LLM categories

EUhosted · GDPR

Next.jsTypeScriptPostgreSQLDockerLLM APIsRBAC

qbitdojo.com

Evaluation infrastructure

Judge-model evaluation pipeline

Cascaded multi-model scoring system that evaluates adversarial prompt defenses across three weighted dimensions: blocks the attack, preserves function, minimality.

Structured outputGraceful degradationToken budgets

Certification methodology

QAISP

Three-tier AI security competency framework (Associate / Practitioner / Expert) with a 150-question exam bank covering prompt injection, model inversion and supply-chain attacks, plus QR-verifiable certificates.

CurriculumAssessmentVerification

Open source · agent opsnpm

FleetHawk

CLI for AI-fleet diagnostics, monitoring and release safety. Surfaces silent failures across Claude Code, Codex, OpenClaw and any framework that leaves config or session artifacts behind.

TypeScriptCLIObservability

github.com/MEdaHad/fleethawk

Autonomous development

MedaXP AI Fleet

Seven-agent autonomous development system: an orchestrator routing work to planning, coding, review, testing, security and documentation agents, with human approval gates on anything irreversible.

Multi-agentTool useEvals

07 off the clock

Side projects, shipped for real

A few personal projects released under the RioticGamer name. They're where I take ideas from work into a different domain: real-time voice AI, computer vision and build tooling, with the same guardrail thinking.

Voice AI · real time

PedTalk AI

Voice-driven, context-aware NPC conversations inside GTA V: speak naturally to characters and get spoken, in-character answers powered by Gemini, with automatic language detection and spatial audio. The professional edition adds multi-person scenes and persistent memory.

29.9k+ viewspublic release · 2026

Release page

Computer vision · native Windows

EYVRO

Webcam head tracking for driving and simulation games, feeding OpenTrack over UDP. A CPU-first pipeline with quick tuning, background operation and camera recovery, built as a local-first native Windows application.

1.0 release candidatelocal-first

Build system · graphics

Halogen

A declarative build system for graphics, lighting, weather, sky and seasons in a truck-simulation game: apply a preset to the game's definitions and get a small mod plus a diff, a report and a manifest of exactly what changed.

declarative presetsin development

08 credentials

Talks, education, languages

Speaking & teaching

Keynote · Swiss Cyber AI Conference 2026 Lugano, April 2026 Hacking LLM Applications: Attack Patterns, Defense Strategies, and Why AI Security Training Matters

Guest instructor · Powercoders ICT Bootcamp Bern, ongoing LLM security and red-teaming sessions for career-changers entering tech

National cybersecurity awareness programmes INSA, Ethiopia · 2018 – 2020 Designed and delivered training for public-sector audiences

Education & certifications

BSc Software EngineeringMicrolink IT College · 2013 – 2017
Cybersecurity CertificationAriel University · 2018
Powercoders ICT BootcampBern · alumnus 2023 · partner & instructor

OWASP LLM Top 10NIST CSFMITRE ATT&CK

Languages

09 contact

Let's build something that holds up.

AI security roles, LLM red-teaming, agent-system design, detection engineering or a talk for your team: I'm based in Lausanne and work with teams across Switzerland and beyond.

Email me Download CV

Emailmedhane.hadush@powercoders.org

LinkedInlinkedin.com/in/medhane-hadush

GitHubgithub.com/MEdaHad