Branch AI

The Problem

• Search queries are becoming more conversational (ChatGPT effect). However, most Ecommerce retailers & brands still use keyword search tools resulting in lost revenue for tail queries.
• Search providers give you tools to build search engines, but leave relevance engineering up to you (i.e. tuning of the search).
• Tuning search to maximize conversions is complicated & expensive (involves stitching together embeddings, rankers, NLP models, heuristics & annotated feedback).
• DIY embeddings and models are easy to experiment with. But building production-grade deployments requires catalog-specific tuning, which can unlock 20-30% better accuracy & conversions.

The Solution

Step 1 - Upload/Connect Catalog: Upload a list of products or connect to an existing search database or platform (e.g. Elastic or Shopify).

Step 2 - LLM-powered Tuning: This happens under the hood automatically. We use LLMs to train catalog-specific embeddings, rankers & query rewriters, bootstrapped from our proprietary Ecommerce-specific models.

Step 3 - AI APIs: With zero effort, businesses get:

1. Search API: Query → Ranked Products (for end-to-end query rewriting, hybrid retrieval & reranking)
2. Query Understanding API: User Query → Catalog-Aware Query (for DIY query expansion/relaxation)
3. Embeddings API: Query or Product → Catalog-specific Embedding (for DIY retrieval)
4. Reranking API: Candidate Products → Re-ordered Products (for DIY precision improvements)

Step 4 - Preferences: Users can control search behavior using:

• Merchandising Strategies: Execute smart strategies using simple instruction templates (e.g. “boost Apple Vision Pros to visitors looking for gifts”)
• Feedback: Demonstrate examples of preferred search results (models learn from these “few-shot”)
• Maintenance Jobs: LLMs periodically evaluate search results (e.g. for queries with zero results) and produce training data to fix gaps

Our Ask

• If you want a better search with lower effort for your Ecommerce store, contact us via email or LinkedIn.
• If you’d like a free evaluation of your site search, submit a request on our website.

About Us

• Govind was previously co-founder of Semantics3 (YC W13), an Ecommerce catalog AI company, from 2011 to 2020.
• Govind & Praveen (founding engineer) built embeddings, rankers, and transformer-based sequence models at Twitter that directly drove 1MM+ DAUs.
• Ishan was CTO at Funding Societies, the largest digital financing Fintech in Southeast Asia, where he built and scaled the technology and business from Seed to Series D.

Our Vision

To serve as an Ecommerce AI platform for brands and retailers. We already see that our models do better than mainstream alternatives at out-of-domain tasks like recommendations, assortment curation, insight extraction (from query text), marketing content generation (from query text), in-session personalization, and event-based ranking.