Portfolio / Projects
Threat intelligence - DFIR - Offensive Security - AI Security

Selected projects

Cybersecurity projects with technical evidence.

Built applications and engineering deliverables are listed here. Controlled security assessments and investigations have dedicated Labs and Writeups pages.

Smart contract security / Academic team project / 2026–2027

CrowdFund DAO — Secure Blockchain Crowdfunding

Three-person Ethereum crowdfunding dApp with ETH contributions, CFD governance voting, multisignature withdrawal controls, IPFS metadata and security testing.

CrowdFund DAO architecture Contributors connect with MetaMask and fund a Solidity crowdfunding contract. Campaign metadata is stored through IPFS, contributors receive CFD governance tokens, and a snapshot-based DAO vote controls fund releases. Security controls protect the flow. CROWDFUND DAO · ON-CHAIN FLOW MetaMask users Create · fund · vote Crowdfunding.sol Campaign lifecycle ETH accounting CFDToken ERC-20 · Votes 1,000 CFD / ETH DAO governance Snapshot vote Release / freeze IPFS metadata Image + JSON CID Security controls Roles · pause · nonReentrant · 2-of-3 above 10 ETH Funds unlock only after campaign success and an executed Release vote
CrowdFund DAO · MetaMask → Solidity → CFD governance → controlled release
  • My work: CFDToken (ERC-20/Permit/Votes), crowdfunding lifecycle, contribution and refund logic, and explicit fund-accounting invariants.
  • Validated refund and reentrancy protections with Hardhat tests and an attacker contract; the team also implemented role-gated pause controls, snapshot DAO governance and 2-of-3 trustee approval for payouts above 10 ETH.
  • Team of three, credited in the case study. This public fork preserves the shared project history and links back to the original team repository.
Solidity · OpenZeppelinHardhat · Security testsERC20Votes · DAOTeam project

AI Security / Red Team Engineering / Project in development

Red Team Agent for LLM Application Auditing

Building a controlled security-auditing toolchain for LLM-powered applications, with repeatable adversarial scenarios, independent evidence checks and explicit limitations.

  • Created UniAssist, a fictional local FastAPI/Ollama test target using synthetic data; recorded initial manual checks of instruction-following and context disclosure.
  • Developed a separate audit engine for scenario validation, HTTP execution, deterministic verdicts and structured reporting on development branches.
  • Preparing an installable Python CLI, optional attacker-LLM follow-ups and a local dashboard; real integration and final review are still pending.
Engineering evidenceScope · Experiments · Validation · Limits

Scope

Local or explicitly authorized LLM applications, not public target discovery.

Experiment

One canary intentionally placed in model context was reproduced; this was not private-file access or production exfiltration.

Validation

Deterministic scenario evaluation and offline tests are implemented in development; broader real-model benchmark remains pending.

Source status

Implementation remains private until technical review and release checks are complete.

LLM SecurityPython · FastAPIAdversarial testingIn progress

Threat intelligence / PFA completed / July-August 2026

CyberVeille PME - suspicious URL detection MVP

Completed PFA project with CMRPI - Espace Maroc Cyberconfiance: an SME-oriented MVP for analyzing suspicious URLs without opening them.

  • Combined local URLHaus lookup, 21 URL features, scikit-learn Logistic Regression, risk levels, reasons and practical recommendations.
  • Implemented SQLite analysis history, CSV bulk analysis, configurable SMTP alerts, optional AbuseIPDB checks and SME reflex sheets.
  • Validated invalid inputs, URLHaus behavior, CSV processing, SQLite storage, disabled alerts and SMTP failure cases with automated tests.
INPUT Manual URL paste + submit CSV Bulk bulk URLs ANALYSIS URLHaus Lookup local database 21 URL Features lexical + host Logistic Regression scikit-learn Risk Levels Reasons + Advice OUTPUT Streamlit UI verdict + details CSV Export annotated results OPS & ALERTS SQLite History SMTP Alerts AbuseIPDB Check Reflex Sheets Auto Tests
Engineering case studyScope · Architecture · Validation · Limits

Challenge

Help an SME evaluate suspicious HTTP(S) URLs without automatically visiting potentially malicious websites.

My contribution

Developed the local URL processing flow, URLhaus lookup, lexical feature extraction, risk classification, Streamlit interface and optional operational alerts.

Validation

The public repository includes automated tests for malformed inputs, missing model assets, bulk CSV processing, history persistence and SMTP error paths.

Honest boundary

Local threat-intelligence data and trained model artifacts are not committed. An unavailable component must be reported as unavailable, not as a fabricated safety verdict.

Threat intelligencePhishing detectionStreamlitSME security

RAG / Secure backend / AI Security

Chatbot USMS academic assistant

Built a multilingual academic assistant for university FAQs, timetables and administrative information, using a RAG architecture and a full-stack application structure.

  • Designed a document-processing pipeline for PDF, JSON, TXT and Markdown sources with chunking and semantic retrieval.
  • Combined FastAPI, Next.js, PostgreSQL, Redis, Qdrant, Docker Compose and SSE streaming.
  • Added security-oriented controls such as JWT, RBAC, rate limiting, audit logs and privacy workflows.
INPUT PDF / JSON / TXT / MD Documents Chunking FastEmbed Embeddings PROCESSING Qdrant Vector Store Semantic search LLaMA 3.3-70B (Groq) Generation + SSE streaming FastAPI Backend PostgreSQL Redis OUTPUT Next.js Frontend Chat UI SSE Streaming SECURITY JWT Auth RBAC PII Masking Audit Logs Rate Limiting
Engineering case studyRetrieval · Security controls · Boundaries

Challenge

Make university reference documents easier to query while preserving access boundaries and providing traceable academic answers.

My contribution

Designed the ingestion and retrieval architecture, application/API separation, and an engineering plan for authentication, access roles, rate limiting and auditability.

Technical decisions

Document chunking and vector search support retrieval; the FastAPI and Next.js architecture separates service logic from the interface.

Honest boundary

The full source repository is not public. The architecture describes the project; request a walkthrough to discuss implementation details and verification.

FastAPIRAGPostgreSQLSecurity

Applied ML / Risk classification

Fault classification prototype from sensor data

Built a Python prototype that classifies pump fault risk from sensor values and presents the result through a simple desktop interface.

  • Prepared data with Pandas and trained a Random Forest classifier with Scikit-learn.
  • Translated model output into a readable risk level for the user.
  • Useful supporting evidence for cybersecurity contexts involving operational data, risk classification and explainable output.
Predictive maintenance: from sensor measurements to risk feedback Sensor readings enter a Pandas preparation step, a scikit-learn Random Forest model predicts the fault or risk result, and the Tkinter desktop interface displays feedback to the user. PREDICTIVE MAINTENANCE · ML WORKFLOW Sensor dataVoltage · flow PreparationPandas Random ForestScikit-learn Fault / riskModel output Desktop UIUser feedback Sensor inputs also include speed, temperature, pressure and power
ONEE · Operational measurements → classification → desktop feedback
Engineering case studyMeasurements · Model · Interface · Limits

Challenge

Transform pump measurements (temperature, flow, pressure, speed, voltage) into an understandable fault-risk indication.

My contribution

Prepared tabular data with Pandas, trained a Random Forest classifier and integrated its output into a Python desktop prototype.

Application evidence

Screenshot of the pump fault prediction desktop prototypeScreenshot available in the portfolio repository.

Honest boundary

Academic and internship prototype; the repository does not establish industrial deployment or production-grade predictive accuracy. Future work includes model evaluation and comparison.

PythonScikit-learnRisk

Cryptography module / Visual encoding

DotCodeApp — Point-based coding

Academic Python/Tkinter application for encoding and decoding messages with a 3×3 matrix of colored dots and a decoding key.

  • Converts messages into visual point matrices.
  • Imports and exports coded data as PNG images or JSON files.
DotCodeApp visual point coding: encode and decode A message and key are used to encode a 3 by 3 colored dot matrix and export PNG or JSON. The encoded file can be imported and decoded with the same key to recover the original message. ENCODING Text message+ encoding key EncodePython / Tkinter 3×3 colored dots ExportPNG / JSON DECODING Import filePNG / JSON Read matrixColored points DecodeWith matching key Recovered textOriginal message
DotCodeApp · Visual encoding and decoding (not a cryptographic security claim)
Engineering case studyEncoding · Formats · Reversibility

Challenge

Represent short text as a human-visible pattern of colored points using an encoding key, then recover the text by decoding it.

My contribution

Built the Tkinter desktop application, visual 3×3 point representation and encoding/decoding workflows.

Outputs

Encoded content can be exported as an image or JSON; the repository documents importing encoded information for decoding.

Honest boundary

Visual coding is not equivalent to a reviewed cryptographic primitive and should not be presented as encryption suitable for protecting sensitive information.

PythonTkinterPillowJSON

Contact

Open to PFE 2027 discussions.

I can walk through these projects for cybersecurity, AI Security, SOC/DFIR, Offensive Security or secure software opportunities.