The One About Spatial & Edge AI
What changes when AI has to work in the physical world?
From making reliable decisions over spatial data to running models with limited compute, bandwidth, and training data, join practitioners and researchers as they share the trade-offs behind building AI for real-world and edge environments.
More About the Sharings
Germayne Ng (Head of Cartography Data Science, GoTo Group) & Joshua Choo (Lead Data Scientist, GoTo Group) will share on "Harnessing VLMs in Production: Estimating and Managing Uncertainty in VLM Outputs"
Ask a Vision Language Model (VLM) how confident it is and it'll give you a number. But unlike the confidence score from a trained classifier, that number is itself a predicted token. So how much can you actually trust it?
Germayne and Joshua will share how GoTo tackles this challenge in Cartography, where VLM outputs help determine what map data can be automatically resolved and what should be sent for further review. They'll explore how VLM confidence scores behave, why they aren't necessarily calibrated for the task at hand, and how multiple signals can be combined into a calibrated probability that can be reliably thresholded in production. (Technical Level: 200)
Alex Low (AIoT Lead, KLASS) will share on "AIOT on the Edge"
What happens when your AI model needs to run with limited compute, memory, bandwidth, and data? Alex will explore the challenges of deploying AI on resource-constrained edge devices, including how compact models can be designed and optimised to run on microcontrollers.
He'll also share how to balance local processing with communication across bandwidth-constrained networks, and how to build effective models for specialised applications where training data is limited. Through real-world examples, explore approaches to data collection, augmentation, and transfer learning, and the trade-offs involved in building AI systems where every resource counts. (Technical Level: 200)
Fangzhou Hong (Co-founder & CTO, Ropedia) will be sharing on “HOMIE Gen 2: Capturing Human Experience for Robot Learning”
Building better Physical AI isn't just about collecting more sensor data. Ropedia's thesis is that robot capabilities can scale with the quantity and quality of captured human experience, what they call the “Experience Scaling Law.”
Fangzhou will share the development journey from HOMIE Gen1 to Gen2, and how the system evolved from capturing data to capturing structured, training-ready human experience. Dive into the engineering behind immersive first-person capture, synchronised multimodal understanding, and training-ready data pipelines, alongside the technical trade-offs involved in designing each. He'll also unpack the thinking behind Experience Scaling and what it could mean for how we build and train the next generation of Physical AI systems. (Technical Level: 100-200)
More About the Speakers
Germayne Ng has more than a decade of experience building high-impact and scalable data science capabilities across ride-hailing, on-demand services, and healthcare. He currently serves as Head of Cartography Data Science at GoTo Group, where he oversees data science work across mapping, including arrival-time estimates, food preparation-time predictions, and pickup-point recommendations. His team also builds and maintains the mapping infrastructure behind these services, covering roads, points of interest, and AI pipelines that keep map data accurate and up to date.
Joshua Choo is a Data Scientist on the Cartography team at GoTo Group, where he works on maintaining accurate map data for routing and arrival estimates behind food orders. Roads change, shops close, and new ones open; much of his work involves ingesting data from multiple sources and modalities, using machine learning and AI to maintain an accurate map at scale. Before moving into mapping, he spent several years as a researcher at the intersection of cybersecurity and machine learning.
Alex Low is a Technical Lead at KLASS Engineering & Solutions Pte Ltd with over 10 years of experience delivering complex AI, embedded systems, and system integration projects from concept through deployment. His work spans engineering execution, system design, and stakeholder alignment, with experience leading multidisciplinary teams, defining MVPs around real-world constraints, and delivering field-ready solutions across complex, multi-stakeholder environments. Notably led the development and deployment of embedded sensing systems for disaster response operations during an international field deployment in Myanmar.
Fangzhou Hong is Co-founder and CTO of Ropedia, dedicating in building the encyclopedia of human experience for embodied AI — the data infrastructure layer that turns real-world experience into machine intelligence. He led the creation of Xperience-10m, Ropedia's flagship multimodal dataset with over 2.7M downloads on Hugging Face, ranking Top 3 on the platform's weekly trending list. Fangzhou holds a Ph.D. from NTU (MMLab@NTU/ S-Lab) and a B.Eng. from Tsinghua University, and is a Google Ph.D. Fellow and China3DV Rising Star. He previously worked with Meta Reality Labs Research.
More About The Series
AI Wednesdays is Lorong AI’s weekly gathering, bringing together practitioners, researchers and innovators for technical discussions on research insights, product development and engineering practices.
Get involved: Learn more about Lorong AI | Speaker Sign-up | WhatsApp Community | LinkedIn | X