About the Role
We're backed by Accel and Weekend Fund and building conversational AI that makes text conversations feel human-like. Our first product, Vaya, is an astrology app where AI personas handle emotionally complex, multi-turn conversations. We've built our own voice model, memory system and evals, and are moving toward fine-tuning our own conversational model.
Turn judgment about what makes a conversation feel right into rubrics, eval datasets and better prompts.
What You'll Do
- Read real user conversations every day. Spot recurring patterns in what the AI says, how it says it, and where people get frustrated
- Turn each pattern into a small, focused dataset of real examples, not one giant test set that goes stale
- Write rubrics so reviewers and automated judges score conversations the way you would, with separate dimensions for sounding human, validating feelings, tracking specifics, and moving the conversation forward. One overall score hides regressions
- Build test sets for the hard conversations: anger, pushback, skepticism. Polite reassurance fails there. Check whether the AI can name the tension, set a boundary, or gently challenge instead of only soothing
- Watch the AI's repertoire. When it leans on a narrow set of moves like reassuring and agreeing, push it toward the range good supporters use: clarifying questions, reframing, honest disagreement
- Iterate on prompts side by side, judge the outputs by hand, and ship the changes yourself
- Run user interviews and studies to check that what you see in logs matches what people actually feel
- Decide what the AI must never do in emotionally sensitive moments, and build the tests that hold that line
What We Look For
- 3+ years in a research role with a record of delivering behavioural user insights
- Experience understanding people in difficult emotional situations, beyond generic empathy
- Ability to spot patterns across conversations and explain precisely why a response works or fails
- Research discipline: distinguish anecdotes from patterns and design tests that settle disagreements
- Clear, defensible judgment when automated scores aren't enough
- Comfort working with prompts, datasets and eval tools. You don't need to be an engineer
- Ability to recognize when a comforting-sounding response could still cause harm
- Background in behavioral science, psychology, counseling, conversation design or another path that built these skills. No specific degree required; what matters is your ability to do the work
How We Work
We care about craft obsessively. Your work gets questioned, pulled apart, and rebuilt — not because we're harsh, but because everyone here holds each other to a standard most places don't bother with. We work out of a hacker house in Vasant Kunj. We strongly encourage everyone to be in office.
Apply
Loading application form…