🎙️ Speech & Audio AI Engineer | ASR · TTS · Voice AI · LLM · Production ML
Passionate about building AI systems that understand, generate, and reason about sound. My engineering practice lives at the intersection of speech recognition, acoustic modelling, and real-time voice AI from embedded hardware to cloud-native conversational agents.
I completed my M.Sc. Data Science with a Thesis GPA of 1.0, dedicating my research to lightweight, noise-robust ASR on battery-powered embedded devices compressing a 1 GB model to <500K parameters using Knowledge Distillation, achieving 75% latency reduction and Macro F1 >0.8 across all real-world noise conditions.
💻 Technical Toolkit
| Category | Skills & Frameworks |
|---|---|
| ASR & Acoustic Modelling | RNN-Transducer · CTC · Whisper (LoRA) · WFST Decoding · MFCC · Mel-Filterbanks |
| TTS & Voice Synthesis | Cartesia TTS · Real-Time Voice Serving · Multi-Speaker Pipelines |
| Real-Time Voice Stack | LiveKit Agents · Deepgram STT · Silero VAD · Multilingual Turn Detection · BVC Noise Cancellation |
| Model Compression | Knowledge Distillation · LoRA Fine-Tuning · ONNX · Quantisation · Latency Optimisation |
| LLM & Agentic AI | LangChain · LangGraph · ReAct · RAG Pipelines · ChromaDB · Claude · OpenAI · Gemini |
| Deep Learning | PyTorch · PyTorch Lightning · HuggingFace Transformers · Scikit-learn · Optuna |
| MLOps & Cloud | Python (Advanced) · Docker · CI/CD · Git · AWS (Lambda · S3 · EC2 · Bedrock · CloudWatch) |
🔬 Highlighted Projects
🎙️ Noise-Robust ASR — Master Thesis @ Cardo Systems Lightweight RNN-T/CTC ASR for embedded devices · Knowledge Distillation · 98% compression · 75% latency reduction · Macro F1 >0.8 across motorcycle / skiing / cycling noise profiles
🤖 LiveKit Multi-Agent Voice AI Real-time conversational agent · Deepgram STT + Cartesia TTS + Silero VAD + GPT-4o-mini · 3-agent architecture with tool orchestration and handoff logic · Deployed in production
🧠 RAG Agentic System LangGraph ReAct agent · ChromaDB vector search · Anthropic Claude · Async FastAPI · Docker
📝 GPT from Scratch Transformer LLM built from first principles · Attention · Tokenisation · Training pipeline · Perplexity evaluation
🏆 1st Place — Facebook Women Hackathon 2021 🌍 Berlin, Germany · English C1 · German B2 · Arabic Native 📩 nancyboukamel12@gmail.com · LinkedIn · GitHub

