Sushanti Kerani
Sushanti Kerani

Sushanti Kerani

Applied AI, Software Engineer @ Adobe · Previously at UC Berkeley

I build AI systems that ask before they assume.

Introduction

I started out as a computer vision engineer and backend developer, back when that was the obvious path into ML. As the field shifted toward NLP, I shifted with it, moving into NLP backend work and MLOps, and from there into machine learning, human-AI simulation, and multi-agent orchestration. That's slowly pulled me toward more systemic and agentic AI, mixing in economics along the way, and more recently into interpretability.

Above all, I love building products. Right now I'm most interested in how agents understand context and get better at improving themselves over time. Day to day, I work on applied AI at Adobe.

Before that, I worked across a handful of startups in India, then moved to Berkeley two years ago for my master's. It's been one of the best decisions I've made. California has given me a lot of nature and a lot of sun, and it turns out those might be two of the best things in life.

Experience

  • Adobe
    • Applied AI, Software EngineerAug 2026 – Present
    • Machine Learning Engineer (internship)May – Nov 2025
  • a21.ai
    • Senior Software Engineer, Product LeadFeb – Jun 2024
  • yellow.ai
    • Software Development EngineerAug 2021 – Dec 2023
  • Dragonfruit AI
    • Member of Technical StaffJun 2020 – Aug 2021
    • Member of Technical Staff (intern)Jan – Jun 2020

Projects

CivicSim

An agent-based simulator of U.S. public opinion. Agents are instantiated from American Community Survey microdata and conditioned on demographic-opinion distributions from the Pew American Trends Panel, aiming for a synthetic electorate whose opinions track the real U.S. adult population at the demographic-cell level. Includes experiments benchmarking demographic alignment between the census and survey corpora and testing how well opinions transfer across geographies. The thesis: realistic policy simulation needs grounding in real population data, not just model scale.

BrainSquared

An open-source, local-first agent harness, one interface for every app you use. It builds a living knowledge base from your tools (Gmail, Slack, Notion, your calendar, GitHub) by reading and writing a single Obsidian markdown vault, and surfaces what needs your attention, drafts replies in your voice, and closes tasks back into markdown. Everything runs locally; there's no cloud middleman.