Senior AI Engineer

About the Role

SELLIT9 runs on one number: what a used device is worth. Two things decide whether we get that number right, and you'll own the AI behind both.

The first is condition. Today a seller tells us what shape their device is in and we verify it by hand when it arrives, which caps how fast we can scale and how confidently we can price. You'll build the computer-vision pipeline that grades a device from photos and feeds that grade straight into pricing.

The second is market context. We're building an append-only, point-in-time-correct dataset of competitor payouts and retail prices right now, specifically so a payout model can be trained on it. The deterministic version ships first: measured gaps against rivals, ranked by exposure. You'll be the engineer who takes it from there to a learned model, graded automatically against the price-move ledger we already run.

This is a build-from-scratch role. There's no ML team to inherit: you'll define the damage taxonomy, the labeling process, the evaluation bar, and the retraining loop.

Read the title as engineer, not researcher. Before the AI specialization, you should be someone who could hold a full-stack engineering seat on this team: you design schemas, write and own production services and APIs, deploy them, and debug them when they break at 2am. The models are one part of a system you own end to end. If your experience is mostly notebooks, papers, or fine-tuning behind someone else's platform team, this isn't the right fit.

A note on AI-native, because it cuts both ways: building with agents daily is non-negotiable here, but AI multiplies engineering skill, it doesn't replace it. You need the depth to review, correct, and own everything an agent writes.

AI Tech Stack

We're in the early stages of building our AI capabilities and we're open to the best tools for the job. Here's what you'll be working across:

• How we build: Claude Code across the whole team, a private library of skills and plugins, and MCP servers for our internal tools. Our trade-in flow ships in the Claude Directory as a public MCP connector, so customers can get a quote and place an order without leaving Claude.
• Data foundation: PostgreSQL with append-only, monthly-partitioned market observations; a price-move ledger that grades every price change against a control group; BigQuery for analysis
• Computer Vision: A modern CV framework (e.g. PyTorch, TensorFlow, or equivalent); object detection or image classification tooling; image preprocessing libraries; dataset management and annotation platforms
• LLM & Generative AI: At least one major LLM API (e.g. OpenAI, Anthropic, Google Gemini, or similar); experience with prompt engineering, structured outputs, and tool use; familiarity with multi-modal models (vision + text)
• AI Orchestration & RAG: An LLM framework (e.g. LangChain, LlamaIndex, or equivalent); vector search or retrieval pipelines for pricing and catalog intelligence
• Cloud & Infrastructure: Cloud ML infrastructure (e.g. GCP Vertex AI, AWS SageMaker, Azure ML, or equivalent); containerized model serving
• Languages: Python (primary for AI/ML work); TypeScript / Node.js for integration with our existing stack
• Evaluation & Delivery: Custom evaluation pipelines; human-review workflows; CI/CD for model deployment

What You'll Do

• Lead our CV-powered condition grading system — build the end-to-end pipeline that turns customer device photos into a condition grade (Flawless / Good / Used) that feeds directly into pricing
  • Train and fine-tune models to detect scratches, cracks, dents, and screen damage on iPhones, MacBooks, iPads, and more
  • Build and manage the labeled training dataset; define the damage taxonomy alongside the product team
  • Deploy inference services and integrate with our NestJS backend and pricing engine
  • Use multi-modal LLMs as a validation layer for low-confidence cases and to generate customer-facing damage explanations
• Take pricing from rules to a model — our market-observation dataset (competitor payouts, retail prices, resale comps) is built point-in-time correct so it can be trained on. Start with measured gaps against rivals, end with a suggested payout that gets graded automatically against real outcomes
• Develop AI agents and automation — orchestrate multi-step workflows for condition assessment, pricing decisions, anomaly detection, and internal operations; the useful ones ship as skills and MCP servers the whole team uses
• Integrate AI into the product — wire inference and LLM features into our Next.js / NestJS stack; design for performance, reliability, and scalability
• Own model quality — build evaluation frameworks, implement human review workflows, and close the retraining loop as labeled data grows
• Collaborate with product and engineering to continuously identify high-impact AI opportunities across the trade-in platform

Apply if you

• Have 5+ years of software engineering experience, with meaningful hands-on time in AI/ML production systems
• Are a software engineer first: you could hold a full-stack seat on this team (production services, APIs, PostgreSQL, cloud deployment), and the AI specialization sits on top of that rather than instead of it
• Are fluent in Python — you've shipped CV models or LLM-powered APIs to production, not just notebooks or research projects
• Have hands-on experience with computer vision: object detection, image classification, or damage/defect detection
• Have worked with one or more major LLM APIs and understand prompt engineering, structured outputs, and multi-modal capabilities
• Understand model evaluation — for CV (precision, recall, IoU) and for LLMs (hallucination, grounding, consistency)
• Can deploy and operate ML services in cloud infrastructure
• Are comfortable with early-stage ambiguity — defining the problem is part of the role
• Work AI-natively, and this is non-negotiable: agentic coding tools (Claude Code or similar) are your daily driver, and you can defend every line an agent writes for you
• Are a strong communicator with experience working on international, English-speaking teams.
• Are fluent in English, written and spoken. This is mandatory: specs, reviews, and daily collaboration all happen in English.

Even better if you have

• Experience in recommerce, e-commerce, or marketplaces, where item condition drives the economics
• Built a labeling and annotation operation from scratch, not just consumed someone else's dataset
• Optimized model serving for cost at volume, or shipped on-device / edge inference
• TypeScript and Node comfort for wiring models into our stack (Python stays the primary language for AI work here)