ARMv7 shellcode, security-hardened firmware, ML inference on constrained hardware. Working at the register level — where security actually lives.
Voice manipulation detection pipeline trained on ASVspoof2019. Three architectures compared — BiLSTM, CNN, Transformer — with LFCC feature extraction, feature caching, and EER-based evaluation.
Position-independent ARMv7 shellcode in Thumb mode — null-byte avoidance via encoding constraints, direct syscall invocation. Three-phase project: local execution complete, reverse shell and trace analysis in progress.
Partial Mbed OS 5 HAL on top of ESP-IDF — Mbed threading primitives mapped to FreeRTOS tasks and semaphores. Used to run a Modbus RTU data logger on ESP32 without native Mbed support. Full CRC-16 implementation.
Neural network inference engine in C99 for ESP32 and resource-constrained MCUs. No heap allocation, no external dependencies — forward pass, tensor ops, and weight loading within fixed SRAM budgets. Targets PyTorch-exported models.
Security-hardened firmware stack for ESP32. Full threat model, verified secure boot chain, encrypted OTA pipeline, MPU isolation regions, and hardware-backed key management — full architecture before a line of application code is written.
Android expense tracker with ML-powered receipt scanning via ML Kit OCR, LLaMA 3.3 70B (Groq API) for structured parsing, and an offline Trie + Levenshtein distance fallback classifier. MVVM, Room ORM, WorkManager.
ML-based IDS for MQTT-connected IoT devices. Building a custom dataset by running 8+ attack scripts against a real ESP32 target — covering spoofing, flooding, and injection vectors over HiveMQ. Dataset generation in progress before model training begins.
M1 student in Intelligence et Sécurité des Objets Connectés at Université Moulay Ismail. Graduated top of my Licence d'Excellence class in the same program.
Nearly 19 billion IoT devices are deployed worldwide — in hospitals, power grids, factories, homes. Most people interact with dozens of them daily without knowing. The security engineering behind them hasn't kept up. That gap is what makes the field worth working in, and why it's still largely unexplored.
My focus is low-level — firmware, protocols, and the hardware-software boundary. That's where the interesting security problems actually live. I prefer going deep over going fast. On the AI side, that means everything from deep learning and generative AI to running inference on hardware that was never supposed to run it.
Headed toward research and industry, with the longer goal of building independently.