- What’s old is new: How can a classic two-tower e-commerce recommendation system be used to make agents smarter? How can a classifier be used to make agents more efficient? The ability to take what’s been done and apply it to a new and ever-changing field is paramount for this role. We’ve launched Tool Recommendations as our first foray into agent–led recommendations, and there is a long list of features we want to create on these (& similar) building blocks.
- Own the models: Own the model pipeline from data ingestion to model publishing lifecycle - including models across embedding, reranking, classification, recognition, and other tasks. Be the leader of a group making agent capabilities, not just API calls, but intelligent and efficient abilities that outpace the competition. Provide insight and experience bringing ideas and practices that should be implemented in a stable but young model pipeline.
- Evaluation Obsession: Be willing to stand by your models because you’ve been provided able data to cover every possible outcome you can. Stand by the strengths and be readily willing to admit the faults so the team can be prepared. Be willing to go row by row in a spreadsheet without pride. That kind of activity is “beneath” you because this space is too new to claim any ego.
- Enterprise ready: Everything you build has to run where our customers run: their cloud, their hardware, and sometimes an air-gapped network. Quantization, export, serving, and packaging are part of the model design from day one, not a deployment step at the end.
Our largest customers run Arcade inside their own walls. A Fortune 100 bank doesn't send its agents' tool calls to someone else's API, and its agents can't wait on a frontier model to pick the right tool out of thousands. So the models behind Arcade's agentic features have to be small, fast, accurate, and shippable into a customer's VPC.
This is just one example of why custom, small models are the unlock to many of Arcade’s future products. You'll report directly to the Head of Engineering and own the models that sit in the runtime path of every agent call we serve.
Our first model for agent recommendation is already built (& patent pending), along with its training pipeline, but it needs to be productionized. You'll own it, decide what the ML stack at Arcade looks like, and set the patterns going forward. You’ll own the build-buy decisions for our stack going forward and have a healthy budget to spend.
This role is about shipping. While we're happy to publish what we learn, delivering the product to customers comes first. If you want six months in a notebook before anything reaches production, this isn't the right role. If you want to ship the model and the writeup in the same quarter, it is!