AhmedEladl

Built an end-to-end production AI voice call system (VAD, STT, LLM, TTS) optimized for real-time conversation latency.

• Built and deployed an end-to-end AI voice call system (VAD, STT, LLM, TTS), optimizing the pipeline to achieve ∼400 - 800ms inference latency for natural conversations.
• Curated and preprocessed Saudi dialect audio data, building a full pipeline for fine-tuning open-source TTS and ASR models for high-fidelity speech synthesis.
• Deployed and maintained large language models in production using GCP for low-latency streaming inference.

Processed and analyzed SEC corporate filings using NLP pipelines for large-scale AI disclosure extraction.

• Processed and analyzed SEC filings using NLP pipelines for large-scale text extraction and preprocessing.
• Built labeled dataset of AI-related corporate disclosures and engineered semantic features for quantitative analysis.
• Applied text mining techniques (TF-IDF, keyword extraction, co-occurrence analysis) to identify disclosure themes and patterns.

Developed multimodal RAG pipelines for unstructured documents and natural-language-to-SQL database agents.

• Built an AI system to extract text from various file formats, normalize output, and persist clean structured data.
• Developed a multimodal Retrieval-Augmented Generation (RAG) pipeline for PDFs and images using OCR for semantic retrieval and context-aware generation.
• Designed and deployed text-to-SQL agents to translate natural language queries into executable database queries with automated chart visualisations.

Built an Arabic speech recognition engine with real-time error detection for mispronounced words and stuttering.

• Built an Arabic ASR system with error detection for mispronounced words and stuttering using TTS/STT pipelines.
• Developed automated monthly analytics reports to monitor speech correction progress.
• Engineered deep learning models for stuttering detection and submitted research findings to Interspeech.

Applied deep convolutional neural networks and wavelet transforms to cardiovascular time-series signals and proteomic data.

• Co-authored ~8 research papers applying AI to medical/physiological data and computer science, presented at international conferences.
• Applied Continuous Wavelet Transforms to blood-pressure time series to generate scalograms, training a CNN that achieved 94% test accuracy.
• Used saliency maps to identify critical frequency peaks for disease risk stratification and early prediabetes detection.