RET-AI-L Ready

Inspiration

We were tired of seeing vague product descriptions filled with phrases like “premium materials” and “luxurious feel” that sound appealing but provide little useful information to someone actually deciding whether to buy the product. This led us to want to help sellers market their products with specific, useful, and accurate information that clearly communicates what the product does, which use cases it supports, and what pain points it solves.

With the rise of AI purchasing agents, this becomes even more important. Unlike human consumers, AI agents need structured, factual information to determine whether a product actually meets a buyer's needs. RET-AI-L Ready helps sellers bridge this gap by making their product listings more informative, relevant, and understandable to AI purchasing agents.

What it does

RET-AI-L Ready is a RAG-powered product-listing optimization application for sellers and catalog managers. It converts an incomplete product listing into a structured Product Passport, evaluates the listing against category requirements, user-query demand, and permitted competitive observations, then interviews the seller to close the highest-impact information gaps.

The application proves value with a before-and-after recommendation test. A simulated AI shopper evaluates the product before the interview, the seller supplies missing specifications or evidence, and the same query runs again against the improved Product Passport.

How we built it

We built RET-AI-L Ready as an end-to-end web application that takes a seller's product listing and evaluates how well it can be understood and recommended by AI purchasing agents.

First, the seller uploads or pastes their product listing. We use an LLM to automatically detect the product category and extract the listing into a structured Product Passport, containing important product facts and attributes.

We then run the Product Passport through our AI-readiness evaluation engine, which looks at factors such as product information, shopper intent, search language, evidence, and consistency. Each product claim is also tracked by its provenance — whether it is verified, seller-declared, AI-inferred, or missing — so that we don't accidentally present AI-generated assumptions as facts.

For missing or weak information, our Seller Coach guides the seller through targeted questions to fill the most important gaps. The updated information is then used to generate an improved Product Passport.

Finally, we simulate AI purchasing-agent queries before and after the improvements. This lets sellers see whether their product becomes more relevant to a shopper's needs and whether the available evidence is sufficient for an AI agent to confidently recommend it.

Technically, we built the application with Next.js, with a separate deterministic evaluation engine, OpenAI models for extraction and querying, and Supabase for persistence. We also built automated tests and end-to-end browser tests to verify the full workflow from listing import through evaluation, seller coaching, simulation, and Product Passport export.

Challenges we ran into

Understanding an unfamiliar problem space

We initially had limited exposure to the AI purchasing agent industry, so we had to research the space extensively to understand how AI agents make purchasing decisions and whether this represented a genuine gap in the market. We looked into Rezolve AI to better understand the current landscape and validate that our product was solving a meaningful problem rather than creating a solution without a clear need.

Building an end-to-end AI application

None of us had much experience building a complete application involving a RAG pipeline, so there was a lot of experimentation and tinkering involved. We had to learn how the different components, e.g. the database, the retrieval system to the AI model, and the application, fit together while debugging issues along the way.

Iterating on the frontend

Our initial frontend did not meet the design standard we wanted, particularly in terms of presentation, so we decided to redo it rather than settle for a mediocre interface. This taught us the importance of iterating quickly and being willing to throw away work when it does not serve the final product.

Balancing ambition with time constraints

We had many ideas for features we could add, but limited time meant we had to distinguish between must-have functionality and nice-to-have features. We focused on getting the core AI-readiness evaluation working first before adding additional features and polish.

Connecting the AI output to something actionable

Another challenge was making sure the AI's analysis was not just a generic score or collection of suggestions. We had to think about what information would actually be useful to sellers and how to turn the analysis into concrete recommendations they could act on when improving their listings.

Accomplishments that we're proud of

Turning an unfamiliar problem into a working product

We started with limited knowledge of the AI purchasing-agent space, but were able to research the industry, identify a genuine gap, and turn that insight into a working end-to-end product within the hackathon.

Building a complete AI pipeline

We're particularly proud that we didn't stop at simply generating better product descriptions. We built a complete workflow, from listing ingestion and structured product extraction, to AI-readiness evaluation, seller coaching, and purchasing-agent simulation that directly addresses the problem we identified.

Making AI recommendations grounded and actionable

We wanted to avoid having the AI simply generate generic suggestions. Our system evaluates specific product information and identifies gaps that sellers can actually address, while distinguishing between verified information, seller-provided claims, and AI-inferred information.

Successfully learning technologies we weren't familiar with

Building an end-to-end application with LLMs, structured data, retrieval, and a full web stack was outside our previous experience. We're proud that we were able to learn and integrate these components into a functional product despite the limited time.

Iterating instead of settling

We weren't satisfied with our initial implementation, particularly the frontend and user experience, so we went back and rebuilt parts of it. We're proud that we were willing to iterate and prioritize the quality of the final product rather than simply stopping once we had something that worked.

What we learned

Start with the problem, not the technology

We learned the importance of understanding and validating the problem before deciding on a solution. Since AI purchasing agents were an unfamiliar space to us, researching the industry helped us refine our idea and ensure that our product was addressing a genuine need.

AI products need more than just a good prompt

We learned that building a useful AI application involves much more than getting an LLM to generate an answer. We had to think about structured data, retrieval, grounding, evaluation, provenance, and how to turn AI outputs into actionable recommendations.

Building end-to-end requires constant iteration

We learned that different components of an application are highly interconnected. Changes to the backend, AI pipeline, or data structure could affect the user experience, so we had to continuously test, discuss, and iterate rather than building each component in isolation.

Good UX is just as important as good technology

We learned that a technically functional product isn't necessarily a good product. Our decision to redo parts of the frontend reinforced the importance of clarity, usability, and presentation, especially when building something that needs to communicate complex AI analysis to users.

Teamwork means leveraging different strengths

Working on a project outside our usual experience taught us how important it is to divide responsibilities based on each person's strengths while still communicating closely. We were able to learn from one another and move faster by combining our different skill sets.

What's next for RET-AI-L Ready

If we were to take RET-AI-L Ready further, our next step would be to validate it with real sellers and real-world product listings. We would want to test whether our AI-readiness evaluation actually correlates with how well AI purchasing agents can understand and recommend a product.

We would also explore expanding our knowledge base across more product categories and industries, improving our evaluation criteria with real purchasing-agent behaviour, and integrating directly with existing e-commerce platforms so sellers can optimize their listings without leaving their workflow.

Ultimately, we see RET-AI-L Ready as a starting point for exploring how product information will need to evolve as AI becomes an increasingly important intermediary between sellers and consumers.