Tejas Agarwal
I'm building Alder (YC F25), an AI employee for accounts receivable. Most companies still collect cash from an aging report and one person's inbox. Alder finds out why each invoice is stuck, fixes what it can, and gets it paid, with a human approving anything a customer would see.
Before Alder I spent six years at Uber building large-scale distributed systems, including the systems that paid more than 5 million drivers every week. Before that I built the first data import pipeline for Oracle Cloud Database.
I started in research. I did my MS in computer science at UW-Madison and worked on parallel and GPU systems at the University of Missouri, the University of Calgary and IIT Madras. That work won the Best Poster award at HPDC and the ACM Student Research Competition at PACT. I still review for systems venues, most recently the EuroSys artifact evaluation committee. (Some people know me as Tejaswi; it's the name on my papers.)
What I'm thinking about
- AI that does the work, not drafts of it. The useful agent in finance doesn't send more reminder emails; it does the investigation an AR hire would do.
- Harnesses over prompts. Approvals, memory, evals and a clear record of every action are what make an agent trustworthy enough to touch money.
- Agents as coworkers. A teammate that says what it did and what it needs from you, in the tools you already use, instead of another dashboard.
- The finance back office as a distributed system. Cash moves across ERPs, portals, banks and inboxes that disagree with each other. That's a reconciliation problem I've seen before.
Writing
Alder blog ยท how AR actually works, and where cash gets stuck
Research
Papers are on Google Scholar and DBLP. Talks include GoLab 2020, on migrating 1,500 internal APIs at Uber.
If you run finance at a company that sends invoices, or you're building agents and want to compare notes, send me a note.