Clarify both sides of the market
A marketplace is not a one-sided recommender. Ask which side is scarce, what actions are reversible, and whether the system optimizes buyer success, seller exposure, platform revenue, or ecosystem health.
The strongest answers explicitly name the tension between one user's relevance and the marketplace's long-term liquidity.
Architecture
- Candidate generation: retrieve candidates from search, embeddings, geography, availability, and business constraints.
- Eligibility filters: remove unavailable, policy-violating, overbooked, or incompatible matches.
- Ranking: predict utility for both sides, not only click probability.
- Constraint layer: apply capacity, fairness, price, location, and diversity constraints.
- Feedback loop: log impressions, contact, acceptance, fulfillment, cancellation, and satisfaction.
Labels and model choices
Labels are delayed and multi-stage. A click may not mean a successful match. Model a funnel: impression to click, click to contact, contact to acceptance, acceptance to fulfillment.
Start with gradient boosted trees or wide-and-deep ranking models. Add embeddings for sparse users/items and two-tower retrieval when the candidate pool becomes large.
Metrics
- conversion and fulfillment rate
- time to match
- buyer and seller retention
- exposure fairness and supply concentration
- cancellation or complaint rate
- latency and candidate coverage
Failure modes
- One-sided optimization: improves buyer clicks while starving new sellers.
- Cold start: new supply gets no exposure. Add exploration budget and content-based priors.
- Strategic behavior: sellers game ranking features. Monitor feature manipulation.
- Delayed truth: successful fulfillment arrives late. Use funnel proxies carefully.
What the architect signal looks like
Close with a two-sided metric dashboard and a launch plan that protects the scarce side of the market.