From Fragmented Inventory to Intelligent Marketplace: A 12-Week Rebuild of a Salvage Parts Platform

# From Fragmented Inventory to Intelligent Marketplace: A 12-Week Rebuild of a Salvage Parts Platform  ## Overview An automotive salvage‑parts marketplace approached us with a clear mandate: accelerate growth without compromising trust. The platform connected salvage yards to repair shops and consumers, but the experience was inconsistent. Inventory was messy, compatibility data was unreliable, and the checkout flow forced buyers to call yards to confirm fitment. The business had strong demand, yet performance was stalling due to operational friction. We were asked to modernize the product in 12 weeks, improve search quality, and increase conversion while preserving the existing partner ecosystem. The plan required both a user‑facing overhaul and a deep systems upgrade—especially around data quality, inventory ingestion, and installation workflows. By the end of the engagement, the platform achieved a step‑change improvement in search relevance, inventory activation, and buyer confidence. This case study outlines how we delivered the rebuild: the initial challenges, the goals, the approach, the implementation details, the results, and the lessons learned. ## Challenge The platform had been operating for several years, and the original architecture reflected early‑stage constraints: a single monolith, limited analytics, and a heavy reliance on manual processes. As the marketplace grew, these constraints began to compound. We found three primary bottlenecks: data quality, workflow trust, and system performance. **1) Data quality and compatibility uncertainty** Inventory data arrived in multiple formats: CSVs, PDFs, and manual form submissions. Part descriptions varied by yard and often contained ambiguous or missing specifications. Buyers frequently saw multiple similar listings without certainty about fitment. This led to high pre‑purchase support volume and abandonment during checkout. **2) Trust gaps in installation workflows** Many buyers opted for installation services, but the workflow relied on phone calls to confirm availability and job completion. There was no reliable location validation or standardized proof of work. Yards and mechanics had different expectations, which caused delays and disputes. **3) Performance and scalability issues** Search and category pages were slow. Heavy database joins and ad‑hoc filters made query latency unpredictable, especially during peak hours. The platform had limited observability, so diagnosing performance issues took days rather than minutes. The business didn’t just need design improvements; it needed new core capabilities to enforce trust and scale reliably. ## Goals We aligned with leadership and stakeholders on concrete outcomes. The goals were split into product, operational, and technical categories: 1. **Improve search relevance and buyer confidence** - Make compatibility clearer and more accurate - Reduce search abandonment 2. **Increase inventory activation and onboarding velocity** - Reduce onboarding time for new yards - Automate inventory normalization 3. **Standardize installation workflows and proof of completion** - Introduce job start validation and photo proof - Reduce disputes and operational overhead 4. **Reduce search latency and improve system reliability** - Cut page load times on core browsing flows - Add observability to diagnose issues fast We defined target metrics for the 12‑week period and for the quarter after launch, then used those metrics to guide prioritization. ## Approach The strategy combined three parallel tracks: product design, data/AI modernization, and platform performance. **Product design track** focused on clarifying fitment, reducing friction in the buying flow, and turning installation into a transparent, verifiable service. We ran quick design sprints, validated ideas with repair shops, and tested wireframes with a subset of yards. **Data/AI modernization** focused on building a reliable inventory ingestion pipeline. We introduced an AI‑assisted parser to normalize part descriptions and extract key metadata, then mapped that data to a compatibility model. **Platform performance** focused on reorganizing the backend into modular services and adding caching and observability. We prioritized performance for high‑traffic routes: search, category, and product detail pages. The three tracks were executed in a coordinated 12‑week timeline: - **Weeks 1–2:** Audit, baseline metrics, and experience mapping - **Weeks 3–5:** New information architecture, search UX, and parser prototype - **Weeks 6–8:** Implementation of ingestion pipeline, installation workflow, and service refactor - **Weeks 9–10:** Performance tuning, QA, and data migration - **Weeks 11–12:** Launch, monitoring, and post‑launch improvements ## Implementation ### 1) Inventory ingestion and normalization We started with the inventory pipeline because it was the root cause of both UX and operational problems. The existing process required each yard to upload a CSV and hope that fields mapped correctly. There was no standard for part names or model compatibility. We implemented a multi‑stage ingestion process: 1. **Format normalization:** All incoming files were converted into a canonical schema with standardized field names and types. 2. **AI‑assisted extraction:** A transformer‑based parser extracted make, model, year ranges, part category, and condition signals from the description text. We tuned the extraction with a small labeled dataset created from the top 200 listings. 3. **Compatibility engine:** Extracted data was reconciled with a compatibility table that mapped parts to vehicle variants. This enabled precise filtering on the buyer side. 4. **Human review loop:** For low‑confidence extractions, we queued items in a review dashboard, allowing yard staff to quickly correct or confirm compatibility. These corrections fed back into the training data. This pipeline reduced ambiguity in listings and dramatically improved buyer trust. It also made it easier for new yards to onboard, since the system corrected many common data issues automatically. ### 2) Search and discovery redesign We redesigned search around clarity and speed. Rather than presenting long lists of similar items, we emphasized compatibility, condition, and warranty. The key changes included: - **Compatibility highlights:** Every listing displayed a clear “Fits” badge with year range and trim level. - **Quality signals:** We surfaced yard ratings, part condition grades, and warranty duration. - **Progressive filters:** Buyers could filter by vehicle details without leaving search, which increased engagement. - **Comparison panel:** A lightweight comparison panel let buyers see differences between similar listings without opening multiple tabs. On the backend, we introduced a dedicated search index with normalized fields and tuned scoring to prioritize verified compatibility and higher‑trust sellers. The result was both faster search and more relevant results. ### 3) Installation workflow standardization Installation was a major revenue driver, but it had been loosely defined. We created a structured flow with clear roles, SLAs, and verification steps: - **Job acceptance SLAs:** Mechanics had a limited window to accept jobs. If they didn’t, the job was rerouted to available partners. - **Geo‑verified job start:** Mechanics could only mark the job “started” when within a defined radius of the job location. - **Photo proof and signatures:** Before and after photos, plus customer signature, became required for job completion. - **Status tracking:** Buyers saw real‑time status updates and ETA notifications. These changes reduced disputes and improved overall trust in the service. They also gave the business more reliable data on mechanic performance. ### 4) Modular service refactor The platform used a monolith that combined inventory, orders, and user management. We did not fully rewrite the backend, but we separated high‑traffic components into services to improve scalability: - **Inventory service:** handled ingestion, normalization, and listing updates - **Search service:** managed indexing and scoring - **Order service:** handled checkout, payment, and installation workflows - **Analytics service:** centralized event tracking and metrics reporting We also added a Redis caching layer for hot queries and implemented request tracing to improve observability. This allowed the team to identify performance bottlenecks in minutes rather than days. ### 5) Analytics and instrumentation To measure impact, we standardized event tracking across key journeys. We tracked: - Search sessions and filter usage - Click‑through rate from search to detail pages - Checkout steps and drop‑off - Installation request acceptance and completion times - Inventory activation rates per yard These events fed into a dashboard visible to product and operations teams, enabling continuous optimization after launch. ## Results The results were visible within two weeks of launch, and the quarter‑over‑quarter improvements exceeded targets. The platform also gained a clearer foundation for future growth initiatives. **Key outcomes:** 1. **Search improvements** - Search latency reduced from 2.8 seconds to 1.1 seconds on average - Search‑to‑detail click‑through rate increased by 36% - Search abandonment decreased by 22% 2. **Inventory activation and onboarding** - New yard onboarding time reduced from 10 days to 3 days - Inventory activation rate increased by 41% - Manual data correction tickets dropped by 58% 3. **Installation workflow quality** - Job acceptance time improved by 47% - Disputes related to installation dropped by 52% - Customer satisfaction (post‑install survey) increased from 3.6 to 4.4 4. **Revenue and conversion** - Overall conversion rate increased by 18% - Average order value increased by 12% - Installation attach rate increased by 24% These improvements were validated through both analytics and qualitative feedback from yards and repair shops. The most impactful change, according to stakeholders, was the improved clarity around compatibility and the trust introduced by standardized installation verification. ## Metrics Snapshot Below is a high‑level snapshot of the major metrics tracked before and after the rollout: - **Search latency:** 2.8s → 1.1s - **Search abandonment:** 38% → 16% - **Click‑through rate:** 21% → 28.5% - **Onboarding time:** 10 days → 3 days - **Inventory activation:** 52% → 73% - **Installation disputes:** 4.2% → 2.0% - **CSAT:** 3.6 → 4.4 - **Conversion rate:** 2.8% → 3.3% These metrics helped leadership justify further investment in AI‑assisted enrichment and new geographic expansion. ## Lessons Learned ### 1) Don’t treat data quality as a backend problem only In marketplaces, data quality is directly tied to trust. Improving ingestion and validation had a bigger impact on conversion than purely cosmetic UI changes. The most successful interventions connected data improvements to user‑visible clarity. ### 2) Trust increases when verification is explicit The introduction of geo‑verified job starts and photo proof reduced disputes quickly. Buyers responded positively because they could see the process, not just a status label. This transparency also motivated mechanics to follow the flow. ### 3) Modularization can be incremental A full rewrite wasn’t necessary. By isolating high‑traffic domains into services, we improved performance without destabilizing the business. This approach allowed continued feature development in parallel. ### 4) Observability changes behavior Once engineers had request tracing and standardized event tracking, troubleshooting became faster and more confident. Teams became proactive rather than reactive, because they could spot anomalies immediately. ### 5) Internal tools matter as much as buyer UX The review dashboard for low‑confidence parsing results was critical. It turned AI into an accelerant rather than a black box. Yard staff had ownership, and the data improved continuously. ## What’s Next With the foundation in place, the company can now pursue higher‑impact initiatives: dynamic pricing, demand forecasting, and automated part recommendations. The compatibility model is robust enough to support predictive matching, and the modular services allow future teams to iterate quickly. The next 6–12 months will focus on expanding into new regions, adding more yard integrations, and refining AI‑driven recommendations. The groundwork laid by this 12‑week modernization gives the business confidence to scale without reintroducing friction. ## Conclusion This project demonstrated that marketplace growth often hinges on trust and clarity, not just marketing spend. By investing in data normalization, compatibility validation, and workflow verification, we turned a fragmented salvage‑parts marketplace into a reliable, scalable platform. The improvements were measurable and immediate, and the new architecture now supports long‑term product evolution. For organizations facing similar challenges—messy data, inconsistent workflows, and brittle architecture—the path forward is to align product design with system reliability. The most valuable user experiences are built on the most dependable foundations.