About the Role
SELLIT9 runs on one number: what a used device is worth. Two models decide whether we get that number right, and you'll own both.
The first is pricing. We're building an append-only, point-in-time-correct dataset of competitor payouts, retail prices, and resale comps, built specifically to be trained on. The deterministic version ships first: measured gaps against rivals, ranked by exposure. You'll take it from those rules to a price predictor, graded automatically against the price-move ledger we already run.
The second is condition and recognition. Today a seller tells us what shape their device is in and we verify it by hand on arrival. You'll build the computer vision that grades condition from photos and videos, and that recognizes what an item actually is: category, model, specs, damage.
This is a data science seat that has grown into production machine learning, not a software engineering seat that touches AI. We're hiring for the progression: analysis, then models, then models running in production with real consequences. You'll define the damage taxonomy, the labeling process, the evaluation bar, and the retraining loop, and you'll partner with the engineering team to ship what you build.
ML Tech Stack
We're early and open to the best tools for the job. You'll work across:
- Data foundation: PostgreSQL with monthly-partitioned market observations, the price-move ledger that grades every price change against a control group, and BigQuery.
- Modeling: Python end to end: pandas, scikit-learn, and a modern deep learning framework (PyTorch, TensorFlow, or equivalent), for tabular and vision models alike.
- Multi-modal: a major LLM API for validation layers and damage explanations: prompt engineering, structured outputs, vision inputs.
- Serving: cloud ML infrastructure (Vertex AI, SageMaker, or equivalent) and containerized model serving, monitored in production.
- How we build: coding agents across the team (Claude Code, Codex, and similar), a private library of skills and plugins, and MCP servers for our internal tools.
What You'll Do
- Build the price predictor: start from measured competitor gaps, end with a suggested payout per item, condition, and channel, graded automatically against real outcomes.
- Lead condition grading end to end: photos and videos in, a Flawless / Good / Used grade out, feeding straight into pricing.
- Build recognition models that identify category, model, and specs from images and video, so intake gets faster and less manual.
- Own the data: define the damage taxonomy, design the labeling operation, and keep the observation dataset model-ready.
- Own model quality: evaluation frameworks, error analysis, human review, and the retraining loop as labeled data grows.
- Use multi-modal LLMs as a validation layer for low-confidence cases and for customer-facing damage explanations.
- Partner with engineering to ship your models into the quote flow and the pricing engine.
Apply if you
- Have 5+ years in data science or machine learning, with the progression we're hiring for: from analysis, to models, to models you shipped and ran in production.
- Are fluent in Python and its ML stack: pandas, scikit-learn, and at least one deep learning framework.
- Have hands-on computer vision experience: image classification, object detection, or damage and defect detection.
- Have real statistical rigor: you design evaluations, know when a metric is lying, and can defend a model's behavior to a skeptical room.
- Have worked with major LLM APIs: prompt engineering, structured outputs, multi-modal inputs.
- Work AI-natively, and this is non-negotiable: agentic coding tools (Claude Code, Codex, or similar) are your daily driver. AI multiplies your craft, it doesn't replace it: you can defend every line an agent writes.
Even better if you have
- Pricing, forecasting, or demand-modeling experience where your model moved real money.
- Video understanding: temporal models, frame sampling, or quality assessment from video.
- Built a labeling and annotation operation from scratch, not just consumed someone else's dataset.
- Experience in recommerce, e-commerce, or marketplaces, where item condition drives the economics.