Callum Macpherson
RSS FeedI'm an AI engineer in London. At Runna , I build the ML systems that make millions of people healthier runners. Before that, I spent time at AWS helping customers ship custom AI — from post-trained LLMs and autonomous agents to computer vision pipelines — across robotics, finance, healthtech, aviation, and more. I've worked with every modality you can throw at a GPU: video, images, audio, text.
I write about the engineering that gets models from notebook to production: retrieval systems, voice agents, evaluation frameworks, and the deployment details nobody warns you about.
My current research asks where zero-shot voice cloning models store a speaker's identity: read the write-up , or see all my research .
Featured
-
Where Does a Cloned Voice Live? Prompts and Speaker Embeddings in Zero-Shot TTS
Updated:Zero-shot TTS systems condition on a speaker twice: an in-context acoustic prompt and a speaker embedding. I switched these routes off in six systems (31,989 synthesised utterances) to measure what each one carries.
-
Capacity Planning Qwen3-TTS on an A10G
Updated:A single-GPU capacity study for streaming Qwen3-TTS: why one backend queues, why the other stutters, what a client buffer can and cannot fix, and what it costs per character.
-
Making My Coding Agents Talk Back
Updated:orca-talkback gives Claude Code agents in Orca a voice: a spoken summary of every reply within a second, local text-to-speech, and barge-in so you can talk over it.
-
How to Build a Low-Latency Streaming Qwen3-TTS Server
Updated:Turning Qwen3-TTS into a deployable AWS streaming service with CUDA graphs and FastAPI WebSockets, then measuring where first-audio latency actually goes.
Recent Posts
-
Continual Learning: Which metrics are important?
Updated:Ranking the most valuable metrics for monitoring deployed ML models, from user feedback to distribution drift.
-
LLM Evals: A General Framework for Custom Evaluations
Updated:A general framework for building rule-based and model-graded evaluations for LLM-based applications.
-
Advanced Retrieval for Retrieval-Augmented Generation
Updated:Query expansion, cross-encoder re-ranking, and embedding adaptors for improving RAG retrieval quality.
-
Malicious LLM Prompt Detection in Python
Updated:Building a malicious prompt detector using traditional ML classifiers in sklearn, trained on Deepset's prompt-injection dataset.