Building an agentic system for subjective product discovery
Programming Languages: Python Natural Language Processing: LangChain, Hugging Face Transformers, spaCy, NLTK, TextBlob Large Language Models (LLMs): OpenAI GPT, Gemini, Llama Backend Development: FastAPI Infrastructure: AWS Deployment & Automation: CI/CD Pipelines (Jenkins, GitHub Actions) Databases and Data Warehouses: PostgreSQL, chromaDB Containerization: Docker, Docker Compose, Kubernetes OS: Unix
Product discovery is rarely objective — and search results rarely lead to confident decisions. We built an agentic AI system that helps users choose by translating subjective requests into structured, ranked recommendations.
The Client
The client was developing an AI-driven product aimed at simplifying product discovery across online marketplaces. The goal was to help users make purchasing decisions without manually browsing listings, comparing specifications or reading dozens of reviews.
The system was designed to operate on top of existing marketplaces, aggregating publicly available data and turning it into clear, actionable recommendations.
The Problem: Product Discovery Is Not Search
Online shopping rarely starts with precise requirements. Most users describe what they want in subjective terms: “a good gaming graphics card,” “a reliable humidifier,” “a comfortable mouse.”
In reality, many people solve this by delegating the task to someone they trust — a friend or family member — who manually researches reviews, prices, delivery times and seller reputation, then suggests one or two solid options.
Marketplaces are optimised for inventory exposure and filtering, not for reasoning and decision-making. They provide data, but leave users to interpret it on their own.
Online shopping rarely starts with precise requirements. Most users describe what they want in subjective terms: “a good gaming graphics card,” “a reliable humidifier,” “a comfortable mouse.”
Why This Problem Is Structurally Complex
Subjective product discovery introduces several layers of complexity:
User intent is ambiguous and context-dependent
Relevant information often lives outside marketplaces (reviews, comparison articles, expert opinions)
Marketplaces actively restrict automated access
Reviews and seller data are incomplete or noisy
Ranking requires balancing multiple factors, not a single metric
Solving this requires more than a simple recommendation model or a conversational interface.
The Solution: An Agentic Product Discovery System
We designed and implemented an end-to-end agentic system that mirrors how a human researcher approaches product selection — but operates automatically and at scale.
The system followed a multi-step workflow:
Intent interpretation
Free-form user queries were translated into structured search objectives.
External knowledge retrieval
The system queried Google to identify trusted review and comparison sources, filtering out ads and low-quality content.
Semantic analysis with LLMs
Long-form articles were ingested and analysed to extract product candidates and qualitative signals.
Marketplace discovery
Identified products were searched across marketplaces. Listings were collected using custom web scraping pipelines.
Review and seller evaluation
User reviews were processed with sentiment analysis to classify feedback as positive, neutral or negative.
Weighted ranking logic
Products were scored using a transparent ranking formula that balanced price, delivery time, seller reliability and review sentiment.
Result synthesis
The system generated a concise summary and returned direct links to the most relevant offers.
The full workflow was executed in under a minute and produced results comparable to manual research.
Engineering Challenges
The implementation required addressing several non-trivial challenges:
Operating under marketplace anti-bot protections
Simulating mobile traffic patterns to reduce blocking
Processing large volumes of unstructured text
Designing ranking logic that remained explainable
Managing latency across a multi-step agentic pipeline
The system was fully implemented and technically operational across the entire discovery flow.
Outcome
The solution automated subjective product discovery by combining external knowledge ingestion, LLM-based reasoning and deterministic ranking logic.
The resulting architecture provides a practical foundation for building AI-driven product discovery systems on top of existing marketplaces.
