A knowledge graph, a digital twin, and reinforcement-learning agents - pointed at one messy problem: how do you keep a supply chain standing when the world keeps rewriting the rules?
Here is a thing that is true about supply chains, and also mildly embarrassing: most of them are optimized for a world that does not exist. The forecast says demand will be about 100 units, so you plan for 100 units, and you hold a little safety stock in case it's 110, and everyone signs off, and then a port closes, or a commodity spikes, or a factory in a country you barely think about has a bad month - and suddenly the average was a lie the whole time. It was always a lie. It's just that the lie only becomes expensive on the bad days.
DeepVu - legally still Vufind Inc., a name it kept from a previous life in computer vision - is a small company in San Ramon, California that has decided the bad days are the whole point. Founded in 2016 by Moataz Rashad and Purdue professor Walid Aref, it builds what it calls AI planning agents: software that doesn't just tell you what will probably happen, but decides what you should do about the things that might. The distinction sounds academic until you're the one holding the inventory.
The pitch, roughly, is this. Forecasting is a prediction problem, and prediction problems have been thoroughly colonized by machine learning. Planning is a decision problem - what to buy, what to hold, what to move, when, and at what cost - and decision problems are harder, messier, and much more valuable to get right. DeepVu's wager is that the second problem is where the money and the resilience actually live, and that the way to attack it is with reinforcement learning running on a simulation of your business that's been fed a steady diet of bad news.
That's a founder quote, so treat it with the appropriate skepticism you'd apply to any founder quote. But the architecture underneath it is at least internally coherent, which is more than you can say for a lot of things that get called AI. There are three pieces, and - in the grand tradition of naming everything after yourself - each one starts with "Vu."
A supply-chain knowledge graph that's continuously fed macroeconomic indicators, commodity prices and global trade metrics. The idea: your ERP knows what you have, but it doesn't read the news. VuGraph does, so the external signals end up inside the plan instead of outside it.
A digital twin that simulates your operations under both normal and shock conditions. You don't test a plan by running it once - you test it by running it a thousand times, most of them ugly, and seeing which decisions still hold up when the day goes wrong.
The decision layer: multi-agent reinforcement learning that acts on the twin to recommend shock-resilient demand and supply plans. Now listed on Microsoft AppSource. It's framed as an assistant to human planners - forecasts they can accept or override, not a black box that fires them.
VuGraph pulls your internal data together with external macro signals - trade flows, commodity moves, port congestion - into one connected graph.
VuSim replays the plan across normal and shock scenarios on a digital twin, so a stockout becomes a thing you rehearse instead of a thing that ambushes you.
VuDecide's agents recommend what to buy, hold or move - balancing cost, service and carbon - and hand the call to a planner who can override it.
DeepVu sells modularly, per use case. The same underlying stack gets pointed at whichever planning problem is currently on fire. Relative emphasis, based on how the company describes its own use cases:
Relative emphasis is illustrative, drawn from DeepVu's stated use cases - not a published benchmark.
Figures approximate, compiled from public company data and third-party profiles. Reported total funding ranges from roughly $1.4M to $3.5M across sources.
Two-plus decades across AI, machine learning and computer vision, with stops that reportedly include Sony Ericsson, Samsung and Stanford. Before DeepVu he built Vufind, a visual-intelligence engine for AR and e-commerce - the company DeepVu grew out of.
Purdue professor with a research background spanning Microsoft Research and Panasonic. The academic ballast behind DeepVu's knowledge-graph and reinforcement-learning approach.
Rashad builds a visual-intelligence engine for augmented reality and e-commerce - the computer-vision roots that still live in DeepVu's legal name.
Rashad and Walid Aref pivot the company toward AI-driven, resilient supply chain planning.
A ~$500K tranche lands, with backers including International Venture Partners, Plug and Play and Berkeley SkyDeck.
DeepVu's shock-resilient demand-planning agent goes live on Microsoft AppSource, alongside a blog series on shock-resilient decisioning.
It would be easy to stop at the architecture and call it a day, but that's not quite the whole picture. A public review by the forecasting vendor Lokad scored DeepVu harshly - not for the ambition, but for the opacity, arguing there's a gap between the AI vocabulary DeepVu deploys and the inspectable math it publishes. That's a fair critique, and a common one for early-stage deep-tech companies: the story is coherent, the mechanism is mostly behind glass.
Worth holding both thoughts at once. DeepVu is a small, research-heavy team taking a genuinely hard swing at decision optimization rather than the easier win of a prettier forecast. It's also a company whose strongest claims are, for now, more asserted than demonstrated in public. Both can be true. The interesting question - the one customers eventually ask every AI vendor - is when the vocabulary becomes proof you can see.
DeepVu doesn't maintain a public YouTube channel we could verify, so rather than link to a video that might not exist, here are the primary sources - the product listing, the blog, and the marketplace page where VuDecide actually lives.
How reinforcement-learning agents on a digital twin choose what to buy, hold and move.
How a visual-intelligence startup, Vufind, became a supply chain planning company.
Why VuGraph folds external macro signals into the plan instead of leaving them outside it.
Examining DeepVu's thesis that resilience and efficiency don't have to trade off.
How a tiny team spreads deep-tech R&D across California, France, Canada and Ireland.
What an AI-heavy planning vendor owes its customers when the critique is transparency.