I build the machinery around AI models that keeps them reliable in production, from schema to ship.
The answer-engine-optimization agent I built inside Surface Labs' growth platform. It runs a set of target prompts through ChatGPT, Gemini, Perplexity, and Claude to measure where the brand shows up and whether Reddit gets cited as a source, ranks the coverage gaps, and closes them by generating grounded Reddit comments across 12 subreddits, no product mentions, no links.
Caduceus is the Go backend for a stroke-rehab telehealth platform. Patients train in a mixed-reality headset, a multimodal model scores each rep from the video, and the result lands in their clinical record while the session is still running. Behind it, a Kafka, Redis, and Cassandra pipeline with idempotent writes, because a duplicate rep is a corrupted chart, not a minor bug.
AI phone agents that take restaurant orders and answer questions over the phone, wired into ElevenLabs, Clover POS, Twilio, and Stripe.
A Mixed-Integer Linear Program in Pyomo that places DAG tasks across federated fog nodes to hit a deadline, paired with a greedy heuristic that holds up at 1,000 tasks where the exact solver stops finishing. Published at IEEE Cloud Summit 2025, Washington, D.C.
I'm Amisha. I study Computer Science and Data Science at UC Berkeley, and I build the parts of AI systems that have to keep working after they ship.
Some of that is research. A Mixed-Integer Linear Program I wrote in Pyomo, scheduling work across federated fog systems, got published at the IEEE Cloud Summit 2025 on an NSF grant. Before that, Berkeley's Cognition and Action Lab, turning noisy human-subject data into pipelines that ran the same way twice.
Most of it is production. At Surface Labs, evaluation that runs on every deploy. With UCSF Health and Stanford Medicine, a clinical backend where a duplicate event corrupts a patient record, not a test row. At Amazon, validation across two million records. At Cognova, voice agents grounded in a live POS.
It's the same instinct in all of it. Work out the guarantee the system actually needs, then build the thing that holds it, and leave it simpler than I found it.
UC Berkeley. B.A. Computer Science and B.A. Data Science, with a Computational Biology emphasis. Expected May 2026.
A running build log of what I am shipping and learning. Roughly weekly.
Let's talk.
Open to founding & product-engineering roles, from schema to ship to the on-call rotation.