Suggestions
Yubin Kim
Applied AI/ML leader | Search & Recommendations Expert
Yubin Kim is an experienced applied AI/ML leader and search/recommendations expert currently serving as the Head of Engineering at Vody, a multimodal GenAI startup.12 Here are some key details about his background and experience:
Professional Experience
Vody:
- Joined Vody in May 2023 as Head of Engineering3
Etsy:
- Senior Engineering Manager (March 2022 - October 2023)
- Led a product ML team of 5 applied science and engineering ICs in the Search Retrieval initiative
- Built real-time, large-scale retrieval systems for search, recommendations, and ads
- Directly generated over $40M in Gross Merchandise Sales over 1.5 years through innovations in graph ML models and Solr1
UPMC Enterprises:
- Director of Technology (June 2019 - March 2022)
- Led a product team of 12 engineering/QA ICs and product/engineering managers
- Built an Elasticsearch-based search engine over 160M+ clinical documents
- Led an applied ML team of 3 ML engineers1
Skills and Expertise
- Machine learning
- Natural language processing
- Information retrieval
- Recommendations
- Building collaborative, multi-disciplinary teams
- Solving problems at scale in e-commerce and healthcare organizations1
Education
- Carnegie Mellon University (specific degree not mentioned)1
Location
Pittsburgh, Pennsylvania, United States1
Yubin Kim has over 4 years of leadership experience in applied AI/ML, with a strong focus on search and recommendation systems. He has a proven track record of driving business impact through innovative ML solutions in both e-commerce and healthcare sectors.12
Highlights
So I just realized that I spent the whole of #SIGIR2024 without up-to-date employment on LinkedIn! Now is a good time as any to announce that I've joined Vody in May.
We are a Gen AI start-up determined to make Gen AI actually useful for e-commerce. We have a lot of ideas, but our first product uses multimodal LLMs to fix missing/incorrect information in product catalogues, enriching product data with accurate, buyer-oriented descriptions extracted from a combination of visual/text modalities. And better data means better search, recs, SEO, etc.!
I attended SIGIR this year to get myself up-to-speed on LLMs (my background is in core information retrieval), especially how they're being used in industry. I was looking for ideas on: multimodal LLMs, training/finetuning tricks, prompting tricks, evaluation, LLMs in e-commerce. Keeping this in mind, here are my top takeaways from SIGIR:
Prompting LLMs for best results can be tricky
- Getting good results with soft prompting is difficult and depends on a good starting state, which often requires internal knowledge about the LLM. And if you have the training data/GPUs, might as well finetune your LLM instead of learning a better soft prompt! (This, of course, changes if you don't have the GPUs or you're using closed-source models.)
- Having said that, trying a bunch of different hard prompt variations and picking the best one is easy. So you should do it.
- When fine-tuning, use varying prompts. Also, if you're supplying a list of in-context information, shuffle the list.
- Chain-of-thought is good. The general concept of breaking down the problem for LLMs is very useful for getting better performance and is used across the board, including in LLM-based evaluation.
- Shout-out to the MSFT folks for including a detailed treatment of their prompting strategy in their paper "Large Language Models can Accurately Predict Searcher Preferences" https://lnkd.in/efxB6-A5
- I learned about DSPy (https://lnkd.in/eAQtmdP3), a prompt optimization framework, and I'm excited to try it!
Using LLMs for synthetic search/recs training data generation improves models
- Many industry folks are seeing great results with training their search models on LLM-generated queries, including Spotify (https://lnkd.in/eWaBA3zE), Apple (https://lnkd.in/eJb-W-VT), Comcast (https://lnkd.in/ek3uycF6).
- Best results seem to be when query generation is guided
- Many recommender systems papers are successfully using LLMs to simulate user behavior to generate more training data
Part 2 and 3 are in comments due to character restriction!
I'll be spending the last day at #SIGIR2024 at the workshop I'm co-organizing with Surya Kallumadi, Tracy Holloway King, Vamsi Salaka, Maarten de Rijke. Come to our SIGIR 2024 Workshop on eCommerce!
