Building language-learning tools and LLM pipelines at Lalia

Problem

Lalia builds language-learning products for a fast-growing SEA client base. Content — storybooks, puzzles, audio games — needed to be produced and retrieved at a scale manual authoring couldn't sustain, while staying searchable and consistent.

My Role

Software Engineer Intern, working across the frontend product surface and the data/LLM pipelines behind it.

Approach

• Product: Developed language-learning tools with React and Supabase, including interactive storybooks, puzzles, audio games, and Chrome extensions.
• Data modeling: Architected relational databases for vector search and large-scale asset retrieval — designing schemas, indexes, and transactional workflows to maintain referential integrity across storage systems.
• LLM pipelines: Built LangChain pipelines to orchestrate LLM workflows for image generation and structured metadata extraction, enabling downstream semantic search and analytics.
• Embeddings: Implemented data pipelines for embedding generation, storage, and similarity search.
• Process: Adopted Test-Driven Development and Scrum methodology, and wrote bash scripts to automate recurring tasks.

Outcome

A content pipeline where generation, metadata extraction, and retrieval are connected end-to-end — new learning content can be produced and made searchable without manual bottlenecks, on a schema designed to hold up under referential integrity constraints as the catalog grows.

What I'd improve next

Add automated evaluation for the LLM-generated content (consistency and quality checks before it reaches the product), and expand test coverage around the embedding pipeline's failure modes.