The Challenge
Processing massive volumes of continuous real-time bids with strict latency limits before modern cloud scaling paradigms existed, while protecting privacy.
The Approach
With my team, I designed a decentralized architecture where continuous ML learnings were compressed into Java objects on individual edge servers for sub-millisecond decisions. We pushed parallel bid requests to the browser via JavaScript with a cryptographic sandbox layer (patents: US 9,886,718 & US 20110173126).
Impact
We scaled the self-serve DSP to 1B+ daily bids, and I established relationships with major clients including DoubleClick (Google), ValueClick (Epsilon), and Specific Media (Viant). The company was acquired by News Corp for ~$140M.
The mechanism my team described in 2005 and put into production in 2007 now clears most of the digital advertising transacted online, an auction that runs billions of times a day between machines with milliseconds to decide. The instinct that made it work, pushing the decision to the edge where the context lives and compressing everything the system has learned into something small enough to act on in real time, is the one I use now on agentic AI systems that have to act on incomplete information without a human waiting in the loop.
