BroutonLab

Prototypes that die in a demo are what this engineering team says it avoids. BroutonLab provides a dedicated team to founders who are building AI products for the industry they know from the inside, and it concentrates on software that is live and used by customers, not slide decks.

Where the work concentrates

The site names four areas:

• Production AI agents: software that plans several steps, uses tools, keeps memory and is tested by evaluation, running around the clock. One example sources candidates on LinkedIn for a recruitment-technology platform, working from loosely worded job descriptions.
• Voice AI: real-time voice agents on Twilio that handle real workflows, for example a bot that phones people on behalf of recruiters.
• Ranking and matching: hybrid search, fine-tuned embeddings, re-ranking and testing for cases where simply calling a hosted model is not enough, for instance matching resumes to vacancies for a staffing-technology start-up.
• Custom machine learning for cases where hosted language models do not fit: tuned models, classifiers, image analysis and forecasting, including a yield forecast built for a farming operator.

Research and engineering depth

Beyond product work, the team has done research-driven engineering. It translated statistical estimation code from R and C++ into CUDA so it could run in parallel on graphics processors, and it built a PyTorch backend for non-standard number formats together with compiler researchers. Earlier projects on the site include a text-to-speech model for a healthcare AR and VR start-up, resume parsing software for an applicant tracking system and a real-time explicit content detection model for an Android app.

How a project begins

A first online session lasts about thirty minutes, during which the team learns the requirements. Data scientists then draft a plan based on current approaches and the client's time and budget, and developers are connected with the client for direct communication and reporting. The team says it builds a first version and iterates until it is ready for production, documents everything, and can provide pipelines so the client can retrain models independently. It also fills AI, machine learning and software staffing gaps, and says it has partnered with AI start-ups for more than five years.