• Interpretability: How do LLMs' multilingual capabilities actually work?
• We kind of know they reason in English, then translate to other languages, but how do they actually learn?
• Data attribution: What types of data address which failure modes?
• Data augmentation: What are other creative methods to augment data beyond using English-based sources/methods?

I believe language models can be useful and valuable by:

• Accomplishing better outcomes with less resources
• Reducing friction to do high value activities so we achieve better outcomes more often

Projects I explored:

• Long document AI summarization when details matter:
  • 1. Patent filing optimization: Intellectual property (IP) lawyers currently manually sieve through IP databases to ensure new patent applications do not infringe on existing patents. Indexing technical documents with detailed summaries paired with similarity search could automate this process.
  • 2. Automating knowledge products like UptoDate : UptoDate helps doctors stay current with the most updated research publications by summarizing medical research completely manually. AI summarization could automate this process with human oversight, and gradually replace the need for human verification when the summarization pipeline is mature.
• Realtime AI voice assistants:
  • Voice interfaces solve a critical 1:1 user interaction problem that text-based chatbots struggled with
  • Most use cases of information gathering (i.e. forms, order taking) and question answering (i.e. customer support, outbound sales) are well suited for voice interfaces.

Gigit AI is a platform that helps WhatsApp Businesses scale personalized customer interactions with AI-generated responses.

We learned that WhatsApp businesses rely heavily on the messaging app to drive their revenue, but there were no tools built for them to scale supervision of their customer interactions. Their go-to solution was to scale their manpower linearly to their demands, but this is costly and unwise when demand fluctuates throughout the year. We built Gigit to offer a flexible solution to meeting our customers' fluctuating needs using AI.

I worked extensively on retrieval augmented generation (RAG) applications, developing our own techniques and systems-based approaches to achieve reliability of our AI responses and experimented with various integrations to serve different user experiences. Learned a lot about considering timing of ideas, understanding technology maturity and being pragmatic about execution risk.

I'll always cherish my time at Gigit. We served great customers and innovated intensely, which stretched me to my best.