Building a Memory for the Physical World
From a digital knowledge platform to an operating brain for real-world assets.
In 2024, I left Zhihu and came to Stanford GSB.
For nearly two decades, I had moved from one job to the next without a real break. A weekend, a new office, a new laptop, a new title. This time, I stopped.
I did not have a clean explanation. I was tired, but more than that, I had begun to wonder how much of my career was deliberate—and how much was momentum.
Stanford gave me the distance to reconsider what I wanted to build. It also put me alongside operators from industries far beyond consumer internet, including my classmate and eventual cofounder Steven.
During that pause, I built my first complete program that ran on a server. It was a small thing. But after years of operating a technology company, I was making software with my own hands.
That experiment connected two parts of my life: what I had learned about knowledge at Zhihu, and what I would begin building at Argn.
From digital knowledge to operating knowledge
I spent six years at Zhihu, one of the largest knowledge platforms in the world, and much longer in China’s technology industry.
On the surface, Zhihu is a place where people ask questions and write answers. Underneath, it is a system for turning fragmented human experience into knowledge that other people can discover, evaluate, and use.
An answer is not useful simply because it has been stored. It becomes useful because it is connected to a question, an author, a body of evidence, a reputation system, and feedback from other people. Context and provenance are what turn content into knowledge.
This shaped how I think about AI. Valuable knowledge rarely begins as a clean row in a database. It begins in conversation, judgment, experience, and disagreement. The difficult work is not only generating an answer. It is capturing knowledge without removing the context that makes it trustworthy.
Before Stanford, my cofounder Steven worked in institutional real estate asset management. He once described managing assets in Eastern Germany from an office in London. He had never visited some of the properties. He did not speak German.
The systems around him could report occupancy, rents, traffic, and budgets. They could tell him what had happened. But when he needed to understand why an asset was underperforming, the truth was scattered across property teams, leasing conversations, spreadsheets, weekly calls, and whoever argued most convincingly on Zoom.
High-stakes decisions were being made with low-resolution context.
There is a useful distinction here. Systems of record are designed to preserve outcomes. Operating knowledge explains why those outcomes occurred and what someone should do next.
Real estate has many systems of record. It has very little operating memory.
The knowledge inside a leasing tour
Consider a 45-minute apartment tour.
A prospective renter may say that the unit feels dark in the afternoon. The second bedroom is too small for working from home. A competing property is offering a better concession. Parking is a concern. The price works, but the move-in date does not.
By the end of the tour, most of this information has disappeared.
The property-management system records that a tour happened. The CRM may contain a short note. Weeks later, an asset manager sees that conversion fell or a floor plan is moving slowly. But the prospect’s explanation is gone.
We started Argn by building an AI companion for leasing teams. With consent, it could capture the tour, help the agent remember what mattered, and prepare a more relevant follow-up.
Our early description of the company was an “AI leasing workforce.” I now think that framing was too linear. It jumped too quickly from conversation to automated action.
Before a system can act well, it needs to remember.
Across many tours, individual comments become operating evidence. You can begin to distinguish whether the problem is the agent, the unit, the price, the concession, the marketing source, or the market. When that evidence is connected with inventory, CRM outcomes, competitor supply, pricing, and prior decisions, the property begins to develop a memory.
The workflow is our first sensor. The larger product is the knowledge system built from it.
The edge cases are the product
Working with real properties changed our understanding again.
At one property, several leasing agents shared the same iPad. Unless we solved identity and attribution correctly, the system could assign a tour—and its coaching feedback—to the wrong person.
At another, an agent stopped the recording when she walked back into the office. The system interpreted the missing final minutes as a weak close, when the real issue was incomplete capture.
A unit mentioned casually during a conversation may not match the name in the inventory system. A useful insight for a regional manager may be inappropriate for a leasing agent. A seemingly harmless observation may require careful fair-housing safeguards. A recommendation without visible evidence will not earn the trust required for someone to act on it.
These might sound like implementation details around the product. Increasingly, I believe they are the product.
AI in the real world requires models, data systems, mobile workflows, identity, permissions, evaluation, reliability, and customer judgment to work together. It must survive shared devices, noisy audio, incomplete data, changing operations, and people who have a real job to do.
A model can produce a convincing answer while the system around it is still wrong. It can identify the wrong person, attach an observation to the wrong unit, omit the evidence, expose information to the wrong audience, or optimize a score distorted by incomplete capture.
Building a reliable AI product therefore requires something less glamorous than intelligence: a disciplined respect for reality.
A system that remembers
Internet platforms made digital activity legible. Search engines organized webpages. Social networks organized people and relationships. Knowledge platforms connected questions, expertise, evidence, and trust.
The physical world remains much less legible. Its most valuable context still appears in conversations, site visits, handoffs, local judgment, and small operational decisions. Much of it is never captured. What is captured is rarely connected to the right asset, person, time, and outcome.
We believe the next wave of AI value will not come only from better models. It will also come from making this invisible operating context usable: capturing what people see and hear, connecting it to the correct physical and operational entities, preserving provenance, and learning from the decisions that follow.
Argn is our attempt to build that system for real estate assets.
The first layer is a living knowledge system that remembers what is happening at each property and preserves the evidence behind it. The second is a decision harness that helps operators decide what to change, tracks what they actually did, and learns from the result.
We are starting with multifamily leasing because tours provide a high-frequency view of demand that owners do not have today. The larger idea extends beyond a tour, a property, and eventually real estate.
Every physical industry has knowledge that disappears between frontline work and senior decisions.
We want to build systems that remember.
Come build it with us
Argn is looking for a founding engineer who wants to own this problem end to end.
This is not a role for someone who wants to add an AI feature to an established product. It is for someone who is energized by the difficult boundary between product, backend systems, data, AI, mobile workflows, reliability, and customer operations.
We do not have all the answers. We do have real users, real operational complexity, and a growing conviction that this problem deserves a new kind of knowledge system.
We care less about whether your resume contains every framework in our stack. We care whether you have personally taken ambiguous, consequential problems from first principles to production, and stayed with them when reality broke the clean design.
If that sounds like your kind of work—or if someone you know is looking for exactly this kind of opportunity—read more about the Founding Engineer — Product, Data & AI Systems role or send me a note.