Founding ML Engineer, Computer Vision (Item Identification) (New York)

About the role

Moe's entire pitch to users rests on one claim: we can identify any item in the world from a photo and price it as accurately as a human expert, instantly. You'll build the model that makes that claim true. This isn't a research exercise — every category you get right becomes a category users trust Moe with real money, and every one you get wrong becomes a support ticket and a churn risk. You'll set the technical direction for identification from day one, with real ownership over architecture, data strategy, and the accuracy bar we hold ourselves to.

What you'll do

• Design and own the computer vision architecture for fine-grained item identification — brand, model, edition, variant — starting from foundation vision models and fine-tuning toward Moe's specific catalog
• Define what accurate enough means per category, and build calibrated confidence scoring so the product can say we're not sure instead of guessing
• Build the feedback loop between model errors and what gets labeled next, in partnership with the labeling lead
• Decide where to invest: broader category coverage vs. deeper accuracy on today's categories
• Own the model serving path from research to production — latency, cost, and reliability at scale
• Represent the identification model's capabilities and limits to the rest of the company, including in investor and customer conversations when needed
• What we're looking for5+ years in applied computer vision, with at least one system shipped to production at meaningful scale
• Hands-on experience with fine-grained/instance-level classification, not just general object detection — you've worked on a problem where close isn't good enough (e.g., telling two similar sneaker colorways or watch references apart)
• Strong fluency in PyTorch or TensorFlow, and experience fine-tuning and deploying vision transformers or CNNs in production
• Experience designing and running evaluation frameworks for vision models — you know how to measure whether a model is actually getting better, not just achieving a lower loss
• Comfortable being the most senior technical voice on a hard, open-ended problem with no existing internal playbook
• Strong written and verbal communication — you'll need to explain technical tradeoffs to non-technical stakeholders, including investors

Nice to have

• Prior work at a resale/marketplace company (StockX, GOAT, Vinted, Rebag, The RealReal) or a visual search company (Pinterest Lens, Google Lens, Syte)
• Experience with active learning or human-in-the-loop labeling pipelines
• Familiarity with deploying models behind low-latency APIs at scale
• Founding or early-stage startup experience, ideally as the first ML hire

What success looks like

• 30 days: fully ramped on the current state of the identification problem; has picked the first category to ship (e.g., sneakers or watches) and defined the accuracy bar for it
• 60 days: v1 identification model is live for that category, with a measured accuracy baseline against a held-out test set
• 90 days: confidence scoring is in place, and there's a clear, prioritized plan for expanding into the next 2–3 categories

Comp: $200K–$260K base + 0.75%–1.5% equity