Undergraduate research under Prof. Naseem Daher at AUB: context-aware, learning-driven autonomous ground vehicles for developing countries. PyTorch models, SUMO traffic simulation, ROS2 integration, and principled reformulation of RL exploration bonuses.
AI assistive app for blind & low-vision users — Amazon Industry Program, Team Amazbytes. I own the social-intent-detection module: real-time gesture and interaction detection with MediaPipe and PyTorch, powering proactive, Arabic-first social narration through voice commands (speak, watch, read, find, who).
Hand-rolled Fourier transform and convolution in MATLAB, no built-in fft or conv allowed, ending in a real voice recording filtered clean and deliberately aliased.
Nexys A7 FPGA design building to a six-state FSM midterm, then bare-metal RISC-V assembly on RED-V: an mtime-driven LED flasher and a bit-banged LCD driver.
25-page feasibility report benchmarking ARM against RISC-V on QEMU and Spike, with a total-cost-of-ownership analysis framed for the Lebanese market.
Extended a 5-stage pipelined RISC-V processor to issue two instructions per cycle in Verilog — 727 lines of RTL, six issue rules, and seven testbench cases written to force each rule to fire.
AI clinical triage pipeline built at the Harvard HSIL Hackathon — a FastAPI backend combined with Gemini vision models to prioritize incoming cases.
Stars of Science R&D concept: a low-cost acoustic leak-localization device tuned for plastic pipes where conventional correlators underperform. On-device ML removes the need for specialist operators — targeting non-revenue water loss in the region.
First place, Student Track — EU-funded AI4Purpose Hackathon (MEDAIGENCY). End-to-end multi-turn stateful chatbot on n8n with Twilio WhatsApp integration, feeding a FastAPI dispatch coordinator I built for the responder side.
Keypad-driven combination safe in Verilog, encoder to FSM to seven-segment display, verified with self-checking testbenches covering all 78 possible single- and dual-key press cases.
Raw logs to actionable intel — built at Lebanon's first Agentic AI & Cybersecurity Hackathon (Technologia × 42 Beirut). Automated log ingestion, IP enrichment, and LLM-generated incident reports for the SOC team, orchestrated on n8n.