TechJam Track 4: Shopping Copilot

Inspiration

I am aiming to go into the AI field for a career, hence this hackathon is very relevant in terms of me needing to use a LLM reranker in the project. At the same time, I also get to learn about recommendation system, which is becoming increasingly important in today's data-driven world.

What it does

I have built a fully offline conversational shopping agent that finds a customer’s hidden target product inside a frozen 50,000-product Amazon catalog within at most 10 dialogue turns. The agent parses each simulated customer message into structured preferences, accumulates them across turns, retrieves a candidate pool through a multi-route in-memory pipeline, ranks it with a hybrid lexical + LLM reranker, and asks one clarifying question per turn.

How we built it

I did initial research on retrieval strategies and reranking strategies. I also enlisted the help of Deepseek v4 to help me with my coding and planning.

Challenges we ran into

As I am doing the project solo, there were no one I could discuss with and I have to make all the hard decisions myself such as which LLM Reranker to use.

The main challenge I faced was me not having a NVIDIA GPU. It greatly restricted me on which LLM Reranker I could use and this also means that I am running the reranker with just a CPU, which takes a LONG time to run. I went through many trials and errors with different LLM Rerankers until I settle with one with the best accuracy-time trade-offs.

Accomplishments that we're proud of

This was my first hackathon, so I am proud that I am able to finish it, especially by myself.

What we learned

I have learned a lot about the strategies that go behind a recommendation system. It is really complicated and there is so much to learn but it is also so interesting about how we can use data to predict what people might like.

This was also my first time doing this recommendation system so I have learned a lot about the strategies involved like the lexical scoring, bm25 or even using LLM Rerankers.

What's next for TechJam Track 4: Shopping Copilot

I would probably like to implement dense retrieval, such as by using MiniLM. I did not have much time to do it, but I think it would help greatly in retrieval and the scoring as well. It would be especially important in the real world because it can help with identifying paraphrased text.

Built With

• fts5
• llm
• python
• qwen3
• sqlite