Want to know if this job is worth applying to?
深圳市, 广东省, 中国
Hybrid
ShopBack serves over 20 million shoppers across 13 markets with cashback, deals, and discovery. We're now building the layer on top: AI-native shopping experiences that personalise, guide, and delight at scale.
This role is for someone who closes the gap between product thinking and search depth — and ships AI-powered search experiences that people genuinely use. Not a PM with search opinions. Not an engineer with product instincts. A builder who operates at the intersection and holds the bar across both.
You'll own search quality end to end — relevance, ranking, query understanding, and the eval loop behind them — in a team that is building AI-first from the ground up.
Frame the problem, not just the solution. You take a messy space — shifting user behaviour, an emerging AI capability, a business bet — and sharpen it into a crisp problem statement with a clear success condition.
Set the quality bar. You define what great looks like for search. You review work, call out the delta between good and great, and hold that standard under deadline pressure.
Prototype to think. You use AI-assisted tools to build working prototypes — including ranking and routing changes tested against real queries — before specs are written. You don't wait for a finished spec before forming a point of view.
Lead the product strategy for your domain. You own the outcome, not just the output. You connect metric movements to specific relevance and ranking decisions, make the trade-offs explicit (relevance vs. latency, relevance vs. paid placement), and change direction when the data doesn't support the hypothesis.
Build the operations and eval layer for search. When an AI feature ships, you own the quality bar — eval sets built from real queries, LLM-as-judge scoring validated against human labels, and failure mode coverage. AI output is unverified until tested. "It looks right" is not a quality check.
Lift the team around you. You actively shape how other builders approach problems — through review, patterns, and shared standards. Leadership is a day-1 expectation at this level.
Model AI-native practice for the team. At this level, your job isn't just to use AI fluently — it's to help the people around you understand where it creates real leverage in their specific work. Not sharing tools. Demonstrating deliberate practice.
Maintain search knowledge as a shared layer. Ranking principles, eval sets, and decision rationale kept in formats that engineers and AI tools can consume. Not just slide decks.
Product and technical depth, not one or the other. Strong product judgment with enough technical depth to prototype and reason about search end-to-end — from query understanding to retrieval and ranking. Can spec requirements in forms that directly unblock engineers: verbal description, production URL reference, query set, scoring rubric, or working prototype. Work holds the line beyond the happy path — edge cases, long-tail queries, and localisation.
Genuine fluency with consumer AI products and the Search domain. Has shipped AI-powered features that real users relied on — search, recommendations, ranking, conversational interfaces, or similar. Hands-on with eval pipelines (ideally LLM-as-judge) and familiar with hybrid retrieval and intent classification. Understands where AI adds real value and where it adds noise; can articulate the principles that distinguish the two.
AI as a primary co-worker, not an occasional shortcut. You direct AI across your daily workflow — query-log analysis, judge-prompt writing, eval harnesses, prototype production, research synthesis — and validate outputs against your own judgment. You run materially faster than a conventional workflow without offloading the judgment layer. This is a baseline expectation for this role, not a differentiator.
Seniority that scales beyond yourself. Independently strategic — doesn't need the problem handed to them. At least one team around them is doing better work because of their involvement. Track record of making the right call in ambiguous, cross-functional situations.
The building/selling test. Can point to a consumer metric that moved because of a search decision they owned — relevance, ranking, or query understanding. Can show work that others built on — an eval set, a pattern, a framing — not just work that shipped.
We have moved past "using AI tools." The team operates in an AI-native workflow: prototypes generated before specs, knowledge maintained as machine-readable layers, and eval sets built for AI-assisted surfaces.
AI output on this team is adversarially evaluated, not accepted. Eval sets are shared assets, not personal checklists.
You will be expected to adopt, model, and evolve these practices — not just follow them.