Agent orchestration
Reliable agent orchestration with human oversight in enterprise workflows.
Branavan Selvasingham
I build reliable AI systems for the real world.
Currently VP, Head of AI & Product Engineering at Zafin, focused on reliable agent orchestration with human oversight.
Core themes
Robotics
Embodied AI experiments with the SO-101 and real-world control loops.
Simulation
Multi-agent worlds for coordination, escalation, and behavior over time.
Spatial computing
VisionOS and AR experiments around 3D interfaces and interaction.
Selected Projects
A few projects that show the range of work I do across robotics, simulation, spatial computing, and product building.
SO-101 Robot Experiments
Embodied AI experiments with the SO-101, focused on perception, control loops, and action in the physical world.
See details
Many Lives
A multi-agent simulation exploring coordination, escalation, and persistent behavior in a shared world.
Built to think through how agents behave over time, not just in isolated turns.
See details
VisionOS / Shape of μ
Spatial computing and AR experiments using VisionOS, SwiftUI, RealityKit, and 3D interfaces.
See details
IBM Achievement
Built large-scale systems used by millions and received formal technical recognition at IBM.
Includes 2022 Outstanding Technical Achievement Award work.
See details
Talks
Talks, panels, and appearances on banking, AI, and product engineering.
Includes Microsoft and Money20/20 sessions.
See details
Nimbus POS
A bootstrapped restaurant point-of-sale platform that reached paying customers, exceeded operating costs, and exited through client/data acquisition.
Built and operated from 2015 to 2018.
See detailsCurrent focus
Right now I'm working on:
- Trusted agent orchestration and human review loops
- Reliable AI workflows in production settings
- Robotics and embodied interaction
- Simulation and multi-agent systems
- Spatial interfaces and AR experiments
How I work
A few things that shape the work I like doing:
- I like working across product, engineering, and applied research instead of treating them as separate lanes.
- I'm usually drawn to systems problems where trust, reliability, and judgment matter as much as raw capability.
- I care about making ambitious technical work legible to the people who have to use it, review it, or depend on it.
- I tend to learn fastest by building, operating, and then tightening something into a version people can actually rely on.
Essays
View all essays
More work
Projects by theme
Grouped around the four themes above, plus some earlier work.
Agent orchestration
Zafin
Current work focused on reliable agent orchestration, human review loops, and production AI workflows in enterprise software.
At Zafin, I lead AI and product engineering with an emphasis on turning model capability into systems that can be trusted in enterprise settings.
The work spans workflow design, review and approval loops, operational controls, and the product engineering required to make agentic systems dependable in production.
Podcast with Microsoft's Head of Banking Strategy
A conversation with Tyler Pichach about AI in financial services and enterprise software.
Money 20/20 Presentation (2023)
Presented at Money20/20 on Microsoft’s AI stage, sharing an early roadmap that we later built out and executed.
Earlier orchestration and production AI
IBM production AI and orchestration
Built and shipped production AI systems used by more than 20 million people, including work recognized with IBM’s 2022 Outstanding Technical Achievement Award.
Large-scale AI systems work at IBM that had to perform reliably in production, not just in demos.
The award recognized conversational AI orchestration work that combined model capability with the surrounding system behavior needed for real use.
Auditable Network of Containers
Published IP on transactions between non-trusting parties using auditable networks.
Robotics
SO-101 Robot (LeRobot)
Robotics experiments with the SO-101, including first movement tests and video.
I was testing how well current models can directly control the #LeRobot SO-101 using tool calls, with only one camera mounted on the gripper for world perception and a goal loop (e.g., "pick up pawn" in this case).
— Branavan Selvasingham (@0_branavan_1) December 21, 2025
Takeaway: These models are rapidly approaching embodied AI… pic.twitter.com/AuYLGOFQgn
A hello world of sorts with #LeRobot SO-101. Kinda late to the party because I burned out a motor as soon as I got it earlier 😅. Took some time to get the replacement and get back into it.
— Branavan Selvasingham (@0_branavan_1) September 21, 2025
Really cool tech. Interested to see how much it can do. pic.twitter.com/qwZ5vX5YB1
Vision-Based Robot Arm Motion
Built vision-based motion control experiments for robot arms.
Robotic Conference and Vision
Built a system that combined robotic vision and conferencing for remote interaction experiments.
Smart Locker
Designed and built a smart locker system for secure storage experiments.
Simulation
Many Lives
Simulation and agent orchestration prototype for coordinating multiple semi-autonomous actors in a shared world.
Many Lives is a simulation and agent orchestration prototype built around a shared world with multiple semi-autonomous actors. The goal is to move beyond single-agent chat interactions and explore how agents persist over time, accumulate context, surface competing pressures, and continue acting inside a living system.
The project focuses on orchestration through attention, escalation, memory, and policy rather than direct turn-by-turn control. It sits at the intersection of simulation design, agent architecture, and human oversight for systems that need to remain legible while multiple threads stay active at once.
Boston Dynamics Spot Simulation
Built simulation environments for Boston Dynamics Spot experiments.
Spatial computing
Shape of μ (VisionOS)
A VisionOS project exploring 3D graphing and spatial interaction.
Graph visualizer: "Shape of μ"
Earlier work on interactive systems
Home Hub (Touch, Voice, Display)
Built a home hub that combined touch, voice, and display in one system.
Understanding the tiny bits (img)
Sound Test with touch-based volume controls (vid)
Custom PCBs made from Shenzhen (img)
Display test on assembled custom PCB (vid)
Another display test on assembled custom PCB (vid)
Full system custom PCB with assembled components (img)
First system PCB batch order from Shenzhen (img)
Dynamic Retail E-Paper Labels
Designed e-paper labels for retail stores with real-time updates.
Smart Surface Projection
Built a projection system for interactive display experiments.
Product building
Nimbus POS
A bootstrapped SaaS platform for small cafes with inventory, streamlined ordering, loyalty, analytics, and support for Canadian tax rules. It reached paying customers, exceeded operating costs, and exited through client and data acquisition.
Built features across inventory management, low-click ordering, Canadian tax support, loyalty programs, analytics, and hardware integrations like wireless printing.
Bootstrapped to paying customers, reached monthly recurring revenue above operating costs, and exited through data and client acquisition by Nown POS.
Video Transcript Analysis
Built a web app (thegreatmind.org) for analyzing video transcripts and extracting useful structure from them.
Education
MBA
Schulich School of Business, York University
Focused on finance and entrepreneurship.
Mech Eng
University of Waterloo
Focused on robotics and aerodynamics. 4th year research on ornithopter / drone.