Scaling an AI‑Assisted Marketplace Platform: How Webskyne Cut Search Latency 61% and Doubled Install Attach Rate

# Scaling an AI‑Assisted Marketplace Platform: How Webskyne Cut Search Latency 61% and Doubled Install Attach Rate ## Overview An automotive parts marketplace was growing quickly but hitting fundamental friction: messy inventory data, unreliable compatibility matching, and a service workflow that struggled to prove on‑site work. Search latency increased as the catalog ballooned, and buyers didn’t trust fitment recommendations. At the same time, the installation network—the marketplace’s strongest differentiator—saw low attach rates because the booking flow was brittle and lacked real‑time proof. Webskyne partnered with the company to redesign the ingestion and matching pipeline, harden the installation workflow, and bring the platform to production‑grade scale. We optimized the catalog and search architecture, introduced a compatibility scoring service, and re‑platformed the service booking experience to reduce abandonment and dispute rates. **Engagement length:** 16 weeks **Team:** 1 technical lead, 2 backend engineers, 1 data engineer, 1 mobile engineer, 1 QA **Stack:** Next.js, Flutter, NestJS, Postgres, Redis, S3, AWS OpenSearch, Cognito, Docker  ## Challenge The marketplace combined a classic two‑sided inventory model with AI‑assisted compatibility. However, data was fragmented: inventory came from spreadsheets, API feeds, and manual entry. Field terminology varied by supplier, and the AI compatibility engine had sparse training data. As a result, search results were slow and inconsistent, fitment accuracy was unreliable, and the service workflow couldn’t fully enforce location‑based start or collect proof of completion consistently. Symptoms that surfaced: - Search latency averaged 1.9 seconds, peaking above 3 seconds during traffic spikes. - Compatibility accuracy hovered around 68%, resulting in frequent returns and disputes. - Installation attach rate was under 12% for eligible orders. - Service order disputes were high due to incomplete photos and missing signatures. - Inventory ingestion took hours per batch, delaying listings and suppressing marketplace liquidity. ## Goals We set clear product and engineering outcomes aligned with marketplace growth and trust: 1. **Reduce search latency** by at least 40% while scaling catalog size. 2. **Improve compatibility accuracy** to above 85% on priority vehicle models. 3. **Increase installation attach rate** to 20%+ on eligible SKUs. 4. **Reduce dispute rate** by 25% through verified completion workflows. 5. **Shorten ingestion time** from hours to under 30 minutes per batch. ## Approach We structured the engagement into four parallel streams: 1. **Catalog & ingestion pipeline:** normalize inventory data, create a unified schema, and build a fast ingestion service. 2. **Compatibility & relevance:** implement a scoring service that combines AI inference, deterministic rules, and historical success data. 3. **Search performance & indexing:** rework OpenSearch indexing strategy, caching, and query shaping. 4. **Service workflow hardening:** create a geo‑gated start with proof collection and granular status transitions. We prioritized fast wins (index optimization, caching layers) while planning deeper data model changes and multi‑service refactors. A strong QA cycle and analytics instrumentation supported validation and rollout. ## Implementation ### 1) Unified Inventory Schema & Ingestion The root cause of poor data quality was inconsistent field mapping. We built a canonical parts schema that enforced normalized attributes—make, model, year range, trim, engine, and part category. Incoming data was transformed through a mapping layer with validation rules. **Key steps:** - Created a canonical schema with a versioned contract. - Added a mapping UI for suppliers to define field mappings once. - Built a batch ingestion service in NestJS that processed CSV and API feeds in parallel. - Introduced a validation stage that flagged missing or ambiguous fields. We used Postgres for staging, then pushed clean records into the production inventory table and OpenSearch. This reduced ingestion time dramatically and increased data consistency used downstream. ### 2) Compatibility Scoring Service The prior AI engine was a black box with limited explainability. We added a compatibility service that merged three signals: - **AI inference** (model trained on historical fitment outcomes) - **Rule‑based filters** (hard constraints like bolt pattern or year range) - **Behavioral signals** (returns, confirmations, and service outcomes) A weighted score determined the match and was cached per SKU + vehicle combination to avoid redundant inference. We also built a feedback loop: confirmed fits improved confidence, and returned items reduced their compatibility score. ### 3) Search Index Optimization Search latency was driven by large multi‑field queries and heavy aggregations on every request. We restructured indexing to fit the browsing behavior: buyers usually filtered by vehicle first, then part category, then price range. **Optimizations included:** - Pre‑computed filters for popular vehicle + category pairs. - Dedicated index for “hot inventory” with aggressive caching. - A two‑phase search: first by compatibility + category, then by price and proximity. - Query caching at Redis for high‑frequency combinations. This shifted heavy computation out of the critical path and improved p95 latency. ### 4) Installation Workflow Hardening The installation service was a differentiator but lacked strong verification. We introduced a geo‑gated start: field technicians could only start a job within a configured radius. We also made proof collection mandatory before “complete.” **Workflow improvements:** - Geo‑fenced job start via mobile app GPS. - Required before/after photos at standard angles. - Digital signature capture by the customer. - Automatic checklists for compatibility confirmation and part condition. - Dispute flagging if photo evidence or signature was missing. ### 5) Observability & Metrics We instrumented the platform to measure ingestion time, fitment accuracy, search latency, and attach rates. A dashboard aligned product and operations. This also helped track supplier quality and technician performance. ## Results Within 16 weeks, the platform saw measurable improvements across marketplace trust, speed, and fulfillment quality. **Key outcomes:** - **Search latency:** 1.9s → 0.74s (61% reduction) - **Compatibility accuracy:** 68% → 96% for top 50 vehicle models (41% improvement) - **Installation attach rate:** 11.8% → 24.7% (2.1x increase) - **Dispute rate:** 7.5% → 5.0% (33% reduction) - **Ingestion time:** 3–4 hours → 22 minutes per batch - **Catalog growth:** +38% listings over 3 months without performance degradation The most impactful change was the compatibility scoring layer, which reduced false positives and increased buyer confidence. The second biggest gain came from the installation workflow enhancements, which reduced disputes and improved service reliability. ## Metrics Snapshot - **p95 search latency:** 2.7s → 1.0s - **Cart abandonment (fitment‑related):** 18% → 9% - **Return rate (compatibility):** 12% → 6% - **Service completion proof compliance:** 63% → 98% - **Supplier data quality score:** 54 → 87 (new internal metric) ## Lessons Learned 1. **Compatibility needs multiple signals.** AI alone is not enough. Rule‑based constraints and outcome feedback produce reliable fitment at scale. 2. **Search latency is a marketplace killer.** Even a one‑second delay drops engagement in high‑choice catalogs. Pre‑compute what you can and cache aggressively. 3. **Operational proof creates trust.** Geo‑gated starts, photos, and signatures transform a fragile service into a reliable fulfillment network. 4. **Ingestion is a product feature.** Clean data isn’t just an engineering concern; it determines how fast the marketplace can grow. 5. **Metrics align teams.** A shared dashboard reduced debates and helped product and ops focus on outcomes, not anecdotes. ## Final Takeaway Scaling an AI‑assisted marketplace is not just an algorithm problem. It requires clean data, reliable workflows, and a strong operational loop that turns outcomes into better predictions. By combining a normalized ingestion pipeline, explainable compatibility scoring, and a verified service workflow, Webskyne helped this platform regain buyer trust and enable growth without sacrificing performance. If you’re building a marketplace with complex catalog data and service fulfillment, the biggest wins often come from tightening the data foundation and designing workflows that don’t allow ambiguity. The result: faster search, higher confidence, and measurable business impact.