Paolo Rosson

Paolo Rosson, what I work onThe AI problems I spend my time on.

Most of my work is at Dext, where I lead Applied AI. The rest is experiments I run in my own time.

AI strategy and leadership

Where AI will pay back in a company, what to build first, and which AI to rent and which to own.

At Dext I started the Applied AI function and built Compass, the agent platform our go-to-market teams use every day.

AI in trading and finance

I spent five years at UBS, latterly as a Director, and led one of the bank’s first production LLM deployments, in electronic trading. I know what a trading desk and a model-risk team need before they trust a model.

Private AI on your own hardware

Open-weight models keep sensitive data on your own machines. The hard part is making them fast enough for daily use.

I tune vLLM and SGLang with quantisation and speculative decoding, and publish the benchmarks so others can rerun them. On one NVIDIA DGX Spark I serve a 27B model at 64 to 78 tokens per second for a single user, and 387 tokens per second in total across 32 parallel streams. In one run the machine processed 92.7 million tokens in six hours for $0.22 of electricity, work priced at $460 on a frontier API.

Write-up: a 27B model for 32 streamsWrite-up: a 125B model on one desktop machine

AI agents and document AI

Agents are worth deploying when you can trust what they approve. I design them with evaluation sets, hard limits on cost and time, and checks against adversarial input.

In a public document-intake challenge, my pipeline made zero catastrophic false approvals. It ran offline on a CPU at 2.76 seconds per document, against a 6-second budget, and resisted hidden-text attacks.

Code and results

Generative imagery

I co-founded MenuPhotoAI, which turns a phone photo of a dish into a menu-ready image in about 30 seconds. The hard part is keeping the real food intact while the light, colour and background change.

I built the image pipeline and test prompt systems head to head before each release.

MenuPhotoAI

AI for hardware and wearables

I made the microphone on the Even Realities G2 glasses work as a normal system microphone on macOS, and cut the time to draw a full-screen image on the lens from 9.4 seconds to 449 milliseconds.

I also keep the G2 developer list on GitHub.

Write-up: EvenMicG2 developer list

Get in touch

If you are working on one of these, I am happy to compare notes.

paolo@paolorosson.com