Harish Shankar

Hello! I recently graduated from UMass Amherst in Computer Science, Pure Math, and Philosophy. I completed my honors thesis under Mingda Qiao. I was previously doing research under Cyrus Cousins.

I am interested in the architecture and design choices of language models and making them good for society. In parallel, I also want to continue working on learning theory and theoretical computer science in general.

I spent most of the past year working on Squire (with Aman). Squire was our vision for AI assistants at work and everyday life. It began with us tinkering with fine-tuning (and also distillation) to make a general-purpose assistant that lived on our laptops. We were targeting small models that we could fine-tune for specific tasks and then swap them for the task at hand (our crude version of agent orchestration). However, we soon realized that local inference was the wrong approach.

We transitioned to turning Squire into the AI assistant for all admin work. You could text, call, or email Squire and it took care of scheduling, mailing, researching, coordinating, and filling out forms. More here.

Squire (which effectively concluded as a harness) shut down in March 2026. Retention was poor, and none of the pivots in front of us were exciting enough to justify forcing it. I'd rather work on interesting tech than grind on something I don't believe in. I learned a lot and know what I would do differently next time.