Sam Coope
AI engineer. 9 years working with autonomous agents. Fond of simple solutions to new problems.
github.com/coopie soundcloud London
Positions
- Working with early-stage, VC-backed companies to build the foundations of their agent infrastructure.
- Educating engineering teams on how to design, build and run agents in production.
- First technical hire at an AI recruiting startup (~$20M raised).
- Built Jack and Jill, the candidate-facing and recruiter-facing agents, scaling to ~300,000 live customer-facing agents in production.
more
- Built a fleet of ~30 internal agents running the company's own operations: devops, monitoring, red-teaming, customer success, and abuse moderation.
- Developed the operational playbook for managing agents as a workforce: personality design, interviewing agents during development, giving agents verifiable work.
- Built an LLM-based agent that builds complex backend workflows through a chat interface.
more
- Architected the backend runtime hosting those workflows, doubling as the agent's own development sandbox.
- Every backend service except the runtime ended up hosted on the platform itself, including the coding agent behind the webapp.
- Built the first fully neural voice agent, before LLMs made it easy.
- Hired and managed a team of 12 engineers.
more
- Led core infrastructure development for text classification and NER systems; researched and productionised a novel approach to conversational NER.
- Researched voice cloning and TTS models on an internal NVIDIA NeMo fork.
- Supervised summer interns, resulting in an ACL 2020 short paper.
- Collaborated with marketing on technical blog posts.
- Researched state-of-the-art NLP (extractive question answering, word embeddings) and moved successful models into production.
more
- Developed fast thresholding algorithms to keep automations highly accurate.
- Taught sales teams the AI platform and workshopped effective pitches.
Earlier: ML consultant at Bitfusion.io (2016–17), benchmarking deep learning frameworks on cloud infrastructure · ML intern at AliveCor, Mountain View (2016), deep learning for audio · Software engineering intern at Bloomberg L.P. (2015), moving a large client-side JavaScript app to a headless Node.js service · Software engineering intern at Autodesk (2014), turning slices of 3D models into SVG.
Research
11 publications and 5 patents in conversational AI. First-author papers marked ★.
Using a pre-trained conversational encoder for NER in low-data settings.
Training a sentence-level encoder on Reddit comments for effective sentence representations in conversational NLP systems.
all publications
Using Yelp and TripAdvisor comments to build a restaurant browsing and booking assistant.
Open-sourcing the tools needed to build massive conversational datasets from social media sites.
Neural clustering without a pre-defined number of clusters, using a variation of the noise-as-targets framework.
Outperforming ensemble-based neural NER models by using multiple independent RNNs in parallel.
Parameter-efficient, performant text labelling models built by sharing parameters in an initial encoder network.
Improving attention-augmented answer selection by incorporating a global view of the question and candidate answer.
Adding a Transformer-style attention mechanism to an RNN-based NER model, improving performance on long-distance dependencies in text.