Ravi Patel, Ph.D.

In 2016, riding the post-AlexNet wave, I moved to London to join Entrepreneur First and co-founded Neo AI. EF was valuable because of the room you're in—incredibly smart and driven people given licence to explore, pair up, and solve impossible problems.

Initially with wires, duct tape, and a Raspberry Pi, I built a quantum random number generator (photodiode shot noise). We got it running, but I decided it wasn't the business problem for me.

Looking for a new problem, grounded in what i knew, I started staring at neural network diagrams and math equations trying to figure out how to build a custom chip. Then it hit me—"ReLU was mathematically equivalent to a step function in a boolean space" (ReLU is the activation function for a neural network "gate"). I could replace a ~40K gate floating point multiply-accumulate with a ~400 gate boolean adder. Using code from a University of Toronto graduate course homework assignment, I hacked together a theory-testing prototype using binary weights, simple XOR training, and a sigmoid output—it worked! I showed the prototype output to a person (who I would later call a close friend), and we decided to pair up and build out the prototype binary neural network compiler for FPGAs and TensorFlow.

Three weeks later, the first BNN paper was published out of Bengio's lab. Being scooped felt crappy, but ultimately it didn't matter and we continued to move forward. Eventually we hit operational headwinds and the release of Google XLA and NVIDIA TensorRT as free native tools to speed up neural networks. It's hard to compete with free, so we made the hard call to shut down. In retrospect, we were young and impatient, inexperienced enough to mistake a setback for a dead end. We gave up too early. This was the bitter lesson—persistence and pain tolerance coupled with cleverness is a universal solver.

Healthcare taught me something else: the hardest technical problems weren't the models, they were the data. Clinical data is deeply unstructured, inconsistent, and full of holes, and language models were a seemingly effective way of handling the data structuring problem. I started experimenting with GPT-2 and GPT-3 for healthcare data normalization. Then the "ChatGPT moment" hit, the possibilities exploded, and I stepped away to recharge and go deeper—building agentic systems, testing limits, and learning a new engineering discipline from the ground up.

Today I'm working on agentic AI systems at Delinea (formerly StrongDM)—multi-agent orchestration for cybersecurity, detecting insider threats and account compromise in real time. The work has pushed me deep into what I'd call high-volume agentic engineering: designing systems where dozens of specialized AI agents coordinate autonomously, process massive event streams, and make decisions under uncertainty. It's the convergence of everything—systems thinking from circuits, ML intuition from medical AI, and the infrastructure discipline from scaling a startup.

I'm writing about what I'm learning (as well as some life lessons for my boys) on my Substack.