Aakash Kasamsetty

My first job was making health insurance data trustworthy. Claims, member and plan data arrived in different shapes and rarely agreed. I built BigQuery models over 12M+ claims records and Airflow workflows that normalized them, so retention analytics finally ran off one source of truth.

Then I put it to work: Tableau dashboards on why members left, by region, plan and segment, that helped raise quarterly retention by 2.07%. Validation checks, access controls and field-level masking of member data were built in from the start, not bolted on later. I also modeled player and match data for the Indian Super League into executive reporting.

I started a master's in IT Management, concentrating in data analytics and AI. Moving from data engineering into AI felt like a natural next step: I'd spent years making data reliable, and now I wanted it to do more.

Alongside my studies, I worked as a data analyst for the Registrar's Office, digging into five years of enrollment and waitlist data to find why students kept getting shut out of courses. I joined enrollment, waitlist, section capacity and classroom data into one reporting model, built public Tableau dashboards on demand and room use, and turned the findings into recommendations on section counts, capacity and time slots. Unmet course demand dropped by 18%.

I joined UpX Labs as its second technical hire; the team is now around eight. Here I build AI products end to end, and each one led to the next. USense taught me retrieval and evaluation. UScribe put AI in front of clinicians, where a wrong answer has real cost. And watching those clinicians struggle with scheduling gave me the idea for Vox.

With AI moving this fast, a product can be overtaken in months. USense was: frontier models and longer context windows closed much of the gap it was built to fill, and it was retired. What didn't go out of date was what I learned building it, about retrieval, evaluation and the problems business users actually have with their data. Real problem solving comes from curiosity, not just requirements.

So I'm rebuilding its core in the open as PlainQuery, and testing it against Spider 2.0, a benchmark built from real enterprise warehouses where even the best published agents still miss many questions. The goal is to find out which of USense's ideas still matter, and to measure each one honestly.