Projects where ( Biology ) ∩ ( AI & ML ), a dash of NLP & LLMs
- 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.
You're probably wondering where's the Bio AI projects, here's one ;)
- Parkinson's Disease Voice Detection
Production-grade ML pipeline for detecting Parkinson's Disease from voice recordings using Wav2Vec2 foundation model features—implements dual approach: fine-tuned neural classifier (Wav2Vec2-base-960h → pooling → classification head with dropout) vs Random Forest on frozen embeddings, trained on 81-sample balanced dataset (HC/PD) with stratified split, comprehensive test suite, CI/CD (GitHub Actions), and Makefile workflow—engineered for small medical datasets with configurable YAML-driven training, W&B logging, and device-agnostic execution (MPS/CUDA/CPU).