Angelos Perivolaropoulos is an ElevenLabs research engineer who leads speech-to-text work on Scribe v2 and real-time transcription for voice agents. His work spans speech-model training, production inference, and the infrastructure needed to make demanding AI systems practical.
He studied software engineering at the University of Glasgow, completing an MSci in 2020. His early projects included a reinforcement-learning scheduler for large-scale IoT networks, alongside work combining speech recognition, text-to-speech, and computer vision. He worked at Skyscanner on software engineering and Kubernetes infrastructure, then at Ondat on product reliability, testing, and continuous delivery, before moving into financial-technology platform engineering. That background in distributed systems and operational reliability informs his subsequent work deploying voice models.
Inspired by Andrej Karpathy’s nanoGPT, Perivolaropoulos created llm-from-scratch, an educational GPT-style implementation that trains without pretrained weights on a laptop or accessible cloud GPU. His AI Engineer Europe workshop uses a small Shakespeare dataset to make tokenization, causal self-attention, training schedules, validation, and generation tangible through a complete working model.
Tokenization determines what models can learn. Character-level representations make small-data experiments feasible but scale poorly; speech introduces harder choices about representing voices, music, and other acoustic signals. Perivolaropoulos treats these decisions as central to model design, not implementation details.
Evaluation matters more than falling training loss. He emphasizes learning-rate warmup, controlled decay, validation against unseen examples, and recognizing overfitting. He attributes much of modern reasoning performance to high-quality post-training data and careful evaluation instead of fundamentally different transformer architectures.
Inference efficiency shapes whether voice AI scales. His analysis of voice-model inference economics examines batching, quantization, speculative decoding, and KV-cache optimization. He cautions that compression methods can damage production quality even when narrow benchmarks suggest otherwise.
Decoding should match the product. Transcription rewards consistent, deterministic outputs; creative text generation benefits from temperature and top-k sampling. For multimodal systems, he describes audio or video encoders producing embeddings that language models can process alongside text.
Build a character-level Shakespeare model in PyTorch, following the path from vocabulary design and causal attention to training, sampling, and the limits of a tiny model.
