| 2025 to now |
Google (YouTube), New York
I am the founding PM on a 3-person team picked to bring AI into YouTube's 1,000+ person go-to-market org, where I conceived and built Project Seed, YouTube's AI system for launch communications.
- Seed runs parallel agents over MCP and A2A on Cloud Spanner and Vertex AI, plus a second pipeline that checks thousands of help articles against live product docs to draft updates when specs change.
- Each prompt extracts one signal as JSON checked against a schema, returns null rather than guess, links every sentence to its source doc, and holds 90% factual accuracy on LLM judge and human golden sets.
- Seed automated a 70+ hour workflow across 6 product areas, cut drafting time 80% (116 minutes to 23 minutes), sped up launches 50 to 75%, and freed 15 to 22 people's work across 16+ major launches (now the org default).
- I started as Seed's only engineer and PM and grew it to 5 engineers and a TPM across 10+ partner teams, earning 2 spot bonuses. I also founded Generative AI @ Google (2,300+ members).
|
| 2024 to 2025 |
Google (YouTube), New York
I was the product manager for Bolero, YouTube's system for setting up partner deals and calculating payouts across YouTube TV, Movies, and Music.
- I drove 2-Touch Approvals across Bolero's partner and territory settings, so every change now needs a second person to approve it after reviewing a full diff. We shipped it with sign-off from Finance, Legal, and Engineering after a territory settings error caused a $1.54M overpayment.
- I turned rate cards and subscriber guarantees for NBCU, Fox, and WBD carriage renewals into an automated settlement engine covering $100M+ a year, and I shipped Connected TV partner settlements across $200M a year in payouts.
- I wrote the plan to automate volume discounts and bundles across 41 partners, projected to prevent $79M in overpayments by 2029.
|
| 2022 to 2024 |
Google (YouTube TV), San Bruno
I was a software engineer on machine learning and recommendations for living room streaming and live sports.
- I trained a random forest on watch history to replace heuristic show tags, lifting next-episode F1 from 0.74 to 0.95 and adding $4M a year in watch time (+0.28%).
- I replaced live user clustering with an offline batch pipeline to save 2.5 engineer years of serving cost, and I wrote peak shaving algorithms for NFL Sunday Ticket that cut candidate load 89% (15% of total peak compute).
|
Nights and
weekends |
Independent founder, New York
I build consumer AI products and autonomous Claude agent workflows across 30+ shipped projects at upneja.ai, focused on software that gets people together in person.
- I built Pull Up (pullup.fm) solo on Next.js and SwiftUI, pairing a Bayesian taste engine across 28 dimensions with a nightly Claude agent pipeline that researches NYC live music and texts users when a match is playing.
- I ship daily with fleets of Claude agents behind custom harnesses and skills, running autonomous build loops up to 69 days with a 786-check adversarial eval gate across models.
|
Projects:
Side projects are built on my own time and hardware, separate from my job at Google.
