Skip to content
panoherbe.com
HAM · ––:––:––Say hi →
◎ Live · Hamburg

Yes, it runs object detection on your cursor. Occupational hazard.

Hamburg / 53.55°N, 9.99°E

Pano Herbe builds computer vision, in the wild.

Head of AI at DARVIS — computer vision & generative AI for healthcare and supply chains. Ex-NASA · ex-Volkswagen · ex-Bain · Hamburg.

Portrait of Pano Herbe
face_01.jpg● live

I build computer-vision and generative-AI systems that help hospitals and warehouses see what they already have — in real time, under real constraints.

AI proposes; the human disposes.

Role
Head of AI
DARVIS · since 2021
Prior
NASA · VW
Research → Industry
Based
Hamburg
GMT+1
§02 Selected Work

Where the rigor came from.

05 / items
04/05
DARVIS · Hamburg

Head of AI

Lead the AI org: computer vision and generative AI delivering real-time visibility — Digital Shelves — for hospitals and warehouses. Reliability, privacy-first inference, deep WMS/ERP and SAP BTP integration. Grew from Project Manager to Head of AI, delivering programs across the US, the Gulf, and Brazil.

Computer VisionGenerative AIHealthcareLogistics
Figure of merit
4 yr
PM → Head of AI
§03 /now

A snapshot of current attention.

◉ Focus
  • 01Leading AI at DARVIS — CV + GenAI for hospitals and warehouses.
  • 02Reliable, privacy-first vision pipelines wired into enterprise WMS/ERP.
  • 03Real-time operational intelligence for high-stakes environments.
◐ Exploring
  • Multimodal reasoning for ambient sensing in complex physical environments.
  • Lightweight, on-prem inference architectures for regulated industries.
◎ Open to
  • Speaking on applied AI in healthcare or supply chain operations.
  • Partnerships on computer vision or generative AI that has to survive a real deployment, not just a demo.
DARVIS uptime
0000d 00h 00m 00s
at DARVIS · since 2021-12-01
§04 Signal · a playground

Click the noise. Find a pulse.

iter 000

In 2020 at NASA I trained deep nets to pull pulsars out of radio noise — a precision-vs-recall game with real scientific stakes. This is a playful sketch of that problem. Drop detection taps; tune the signal.

mode = boxes · detections = 0
SNR / dB
−24-12.0+06
Noise floor
clean0.35chaos
this is a toy. the real thing runs on terabytes of radio-telescope data, not random() noise. the instinct’s the same: find the signal, ignore the rest.
§05 Background

Trained in physics, deployed in reality.

MBA
Collège des Ingénieurs · Paris
MSc
Physics · Tübingen
BSc
Physics · Tübingen
§06 Speaking & Teaching

I do the work. I’ll also stand up and explain it.

Teaching
External Lecturer — ICT in Logistics
University of Europe for Applied Sciences
2024 – 2025
Speaks on
01Computer vision for hospital & warehouse operations
02Generative AI on top of vision data
03Shipping reliable, privacy-first ML to production
04From physics to applied AI leadership
Open to conference talks, panels, and guest lectures on applied CV / GenAI in healthcare and supply-chain operations.
Invite me to speak →
§07 Get in touch

Say hello. I read everything, reply to most.