AI-supported admin system for scalable product catalog management Summary

Summary

Boatpartfinder.ai is building a searchable product and spare parts catalog for boat components. The platform collects product data from boat manufacturers, manuals, catalog pages, and other sources. The goal of the thesis is to investigate and prototype the next generation of Boatpartfinder.ai's proprietary CMS/admin system and its features for reviewing, improving, and automating the processing of scraped product data.

The solution should be designed for at least 100,000 products, multiple boat manufacturers, numerous scraping operations (autonomous agents and scripts), and ongoing AI support, where fixed scripts and processes can be gradually replaced as AI continues to improve. The focus is on how humans and AI can collaborate to reduce manual quality work without losing control over sources, traceability, and data quality.

Background

Boatpartfinder.ai is developing a catalog platform for marine spare parts and components. Users should be able to find the right part through search, category, brand, product data, and eventually also by boat model, manual references, and compatibility.

The technical platform currently consists of a public catalog, an internal admin interface, a FastAPI backend, a Supabase/PostgreSQL database, and proprietary scripts and agentic AI workflows. The scraping operations are built on templates for boat manufacturer catalogs: navigation templates are used to find product URLs, and product templates are used to extract product fields from product pages, manuals, and structured data.

Admin features already exist for product lists, categories, brands, duplicate candidates, product verification, scraper templates, scraper status, and AI cost tracking. At the same time, the work is still too manual. An operator often needs to understand why a product was incorrectly scraped, whether a source is reliable, whether a product should be re-scraped, whether two products are duplicates, or whether AI suggestions can be approved.

As the catalog grows from thousands to tens of thousands and toward at least 100,000 products, displaying large lists in the admin is no longer sufficient. The system needs work queues, quality metrics, prioritization, batch workflows, source tracking, AI steps, and clear rules for when a human must review data — as well as a well-developed plan for easily replacing manual steps with AI workflows when the time is right.

Problem Statement

Scraped product data is rarely perfect. A product page may be missing fields, use different product names than the boat manufacturer, mix model numbers with part numbers, contain multiple product variants, link to PDF manuals, or change over time. A conventional CMS is typically built for manually created articles and pages — not for uncertain, semi-automated, and source-tracked product data.

Boatpartfinder.ai has therefore ruled out purchasing a generic or headless CMS. The question is instead how the proprietary admin and CMS layer should be further developed, rebuilt, or built anew on top of the existing database, scraper architecture, and product quality process. 

A well-designed system needs to be able to answer questions such as:

• Which products are missing critical fields?
• Which products are likely duplicates?
• Which products need to be verified against a boat manufacturer source?
• Which fields come from which source?
• Which scraper templates are performing worse than before?
• Which products should be re-scraped?
• Which AI suggestions are reliable enough for automatic handling, and which require human review?

Purpose

The purpose of the thesis is to develop a well-founded vision for a scalable, AI-supported, and proprietary CMS/admin system for Boatpartfinder.ai's product catalog.

The work will combine system analysis, data modeling, UX for internal workflows, and practical prototyping. The goal is to demonstrate how product curation, re-scraping, duplicate handling, and quality control can be made simpler, faster, and more automated as the catalog grows to at least
100,000 products.

Possible Research Questions

• How should a CMS for scraped, uncertain, and source-tracked product data be modeled?
• What workflows are required to manage product quality at a scale of at least 100,000
  products, many of which are very similar to one another?
• Which parts of the current proprietary admin system should be further developed, rebuilt, or
  replaced by new proprietary modules?
• How can AI be used to reduce manual work without compromising reliability and traceability?
• How should human-in-the-loop workflows be designed for product verification, duplicates,
  categorization, and re-scraping?
• How can a system be built so that new AI steps can be added over time without the architecture becoming difficult to maintain?

Proposed Implementation

1. Current State Analysis The student maps the existing platform: the database's key product and source tables, the admin interfaces, the scraper workflow, product verification, the duplicate flow, and
how AI usage is measured. The analysis should identify which parts work well today and which parts will become problematic as data volume grows.

2. Architecture Evaluation for the Proprietary System The student compares different proprietary paths forward. This may involve further developing the current admin, extracting certain parts into new internal modules, rebuilding the operator interface around quality queues, or creating a new workflow layer on top of the existing database and scraper flows.

3. Workflow Design The student produces a proposal for how operators should work with data quality in practice. This may include quality queues, filtering, server-side pagination, bulk actions, source comparison, status models, audit logs, re-scraping, and dashboard views per boat
manufacturer, category, or scraper run.

4. Prototype or Proof of Concept The work should conclude with a prototype demonstrating a central workflow. Examples include a product quality queue, a verification view with source tracking, an AI-assisted duplicate flow, a re-scraping view, or a dashboard for scraper and data quality

Key Design Requirements

Area - Requirement

Scalable search and filtering - Server-side pagination, indexed filters, sorting, and fast views for large datasets

Product quality - Statuses, quality metrics, missing fields, source conflicts, and prioritized work queues

Provenance - Tracking which fields originate from the boat manufacturer, manual, scraper, AI, or human edits

Re-scraping - Ability to re-scrape a product, category, template, or boat manufacturer with clear history

AI support - Modular AI steps for suggestions, classification, summarization,
duplicates, and quality assessment

Human-in-the-loop - Clear boundaries between AI suggestions, automatic changes, and human approval

Operations and cost - Visibility into runs, errors, API costs, batch sizes, and risk of costly operations

Revision - Audit log, field history, and the ability to understand why a product looks the way it does

Suitable Student Profile

This thesis is suited for students in computer engineering, software development, information systems, AI, or database design who aspire to become developers.

Selected candidates are expected to be able to program hands-on themselves, without relying on AI. At the same time, the project is closely tied to modern AI-assisted development, so candidates should also be able to use AI tools, code agents, and so-called vibe-coding in an effective and responsible manner.

Scope Limitations

The thesis does not need to build a complete replacement CMS for all of Boatpartfinder.ai. What matters is analyzing the whole picture and then prototyping a well-chosen part that points in the right direction.

The project should not focus on e-commerce, payments, inventory management, or order workflows. Boatpartfinder.ai is a product and spare parts catalog and is not a retailer in any way.