Projects where ( Biology ) ∩ ( AI & ML ), a dash of NLP & LLMs
- Agents for Computational Biology
From Text to 3D Structure. An intelligent agent powered by Gemini that transforms natural language descriptions of proteins into ready-to-use 3D structures using Apple's SimpleFold. It reasons through biological requirements to generate accurate protein models on demand.
- Arabic NLP Benchmarks & Evaluation
Expanded upon Pico-voice ASR benchmark to support benchmarking of OpenAI's Whisper in transcription of Egyptian ARZ speech into ARZ and english text.
Expanded lm-evaluation-harness to support machine translation of ARZ to english and vice versa with a new ARZ corpus dataset from hugging face and Meta's Nllb200.
- LLM Hallucination Detection & Mitigation
Competition solution for Russian hallucination detection (codeforces) using Agentic Context Engine (ACE)—trained Gemma-3-270M (Generator) with Gemini Flash 2.5 (Reflector/Curator) to learn 83 anti-hallucination strategies from 380 SberQuAD-derived examples that were transformed using Gemini Flash 2.5, achieving 0% hallucination rate with conservative refusal behavior deployed via llama.cpp + GGUF for GPU inference for the competition.
- Geospaital AI Esri North Africa
As part of my GIS intership at Esri, A fine tuned Presto (GeoAI foundational model) on the task of classifying 3 crop types across Africa and also as part of a GeoAI competition on Zindi where it achieved leaderboard position 37, The model's outputs integrate seamlessly into ArcGIs and can be viewed as a feature layer.
- Explainable AI and Adversarial Attacks
XAI, Adversarial attacks and model hardening on CNNs tutorial
A 4 step workflow outlining how to fine-tune ResNet-34/MobileNetV2 on Caltech-101, generate FGSM adversaries using torchattacks, interpret with Grad-CAM & saliency, and evaluate defenses (adversarial training, input transforms).
- Robotics & Voice Control
SHATO — Voice-Controlled Robotic Assistant
A production-grade microservices system that turns speech into schema-validated robot commands—Whisper STT → Fine tuned Gemma3-270M intent extraction → Pydantic validation (self-correction) → Parler TTS—built with FastAPI and Docker, with zero malformed commands reaching hardware.