
Co-founder & CTO @smplrspace · Building The Mesh, a substrate for AI-native teams · Notes at https://nitter.cf/t.co/VP6j6Du1Ua. Scrappy perfectionist, wannabe essentialist.
Singapore
Joined March 2012
- Tweets2K
- Following1.6K
- Followers301
- Likes4.9K
Pinned Tweet
If you're building a crafted product that uses AI to generate work on someone's behalf, you've probably seen it: what it produces doesn't meet the bar the rest of the product sets. A model drifts toward the generic middle of its training data, and that's not where your product lives.
What closed the gap for me wasn't better prompts. It was giving the model my judgment, not just my topic: the standards, the calls I'd make, the things I'd never ship. I've come to think that context, built up, compounds across a team, and that the tools knowledge workers use should help people invest in it, not promise they can skip it.
New post on how I got here, and the question I can't answer yet: whether it holds at a company of hundreds with a product in millions of hands.
t15n.io/where-taste-drops-ou…
Day 17 of the #30DayMapChallenge is light and dark. I recently chatted with a user about a dark mode on the map, so this was a good excuse to try it: here is the southern tip of Manhattan going from day to night, in one slow turn that loops.
I tried two ways to get the dark version. The first was a dark color filter over the whole map, and it came out flat: buildings are all the same white, so a filter can make every one of them lighter or darker but can't give any wall its own shade. The second was to change the light. Move the sun and each wall picks up its own shade, shadows fall across the streets, and once it's dark some of the towers light up.
The map is the @smplrspace map viewer, which runs on @Mapbox. We chose to build on Mapbox for its excellent developer and user experience, and this challenge is a good reminder of why: it made this whole experiment possible. The sun's direction and shadows are set by hand through the day, and Mapbox's own dusk and night lighting take over once the sun is down. Claude wrote the code that renders it frame by frame.
1
8
Day 16 of the #30DayMapChallenge is the collaborative map. Claude can't be my collaborator for this one, so I made mine with my colleague Andrew, and the idea is his.
When Andrew was a kid in Fairfax, Virginia, they said that if you dug through the Earth you'd come out in China. We said the same thing in France, so China must be hard to miss 😅. Andrew wanted to know what is actually on the other side of any spot, and to click back and forth between the two. So that's the map: type an address or click anywhere, press dig, and you get the point exactly opposite, with a map and a street view on each side.
The saying doesn't hold up. Dig from Fairfax and you come out in the Indian Ocean, about 1,200 km from Australia. To reach China you'd dig from Argentina or Chile: the opposite side of the Bund in Shanghai is a country road in Entre Ríos, Argentina.
Most of the time the other side is water, and that was Andrew's idea too: a castaway on a raft, with signposts to the nearest land and how long you'd drift to get there. 🛟
Claude wrote the app. Andrew came up with it, added the street views and the castaway, and shaped the result by trying it and telling me what to change. I set up the project, and did the art direction, which is meant to be great for kids.
You (and your kids) can try it with your own address at dig.smplrspace.io. It doesn't store addresses.
Maps and street views: Google. Photos: Wikipedia and Wikimedia Commons contributors. Land outlines: Natural Earth.
1
100
Day 15 of the #30DayMapChallenge is Inside out, upside down, with one thing on the map not quite right. A mall is a street folded inside a building: shopfronts along a walk, no cars, a roof on top. I did the reverse and drew a real district of Venice as the floor plan of a mall.
The video opens inside what looks like a shopping mall: glass shopfronts on both sides, an atrium, an anchor store, a moving walkway down the middle. Then it rises, the plan turns into a map, and it is San Polo, one of the six districts of Venice. The anchor store is the Frari basilica. The moving walkway is the Rio San Polo, a canal.
Claude did most of the legwork, and nearly all of it through the @smplrspace platform: pulling San Polo's buildings and canals from OpenStreetMap, tracing the district's outline from the map itself, generating the mall as a space with glass shopfronts, and recording the camera. I picked Venice after exploring a few cities, because its canals already read as the corridors of a floor plan I also did the art direction, the cut along the natural paths of the map and a dark floor for the canals.
The mall is synthetic, built from OpenStreetMap data at half size, so its alleys are narrower than they should be.
Data: OpenStreetMap contributors and Mapbox.
1
1
125
Day 14 of the #30DayMapChallenge is the Borgesian map, after Borges's story of cartographers who drew a map of the empire as large as the empire itself, and found nobody could use it. It's a theme about scale. Every map picks one, and the scale decides how much detail makes it in, how much is enough, and when more detail stops helping and starts getting in the way.
I'm late again, so I folded in yesterday's theme, Interactions, which for me goes hand in hand with scale. Detail is worth having when you can see how it connects to the patterns above it. A few months ago, on our podcast Spatial sessions with Andrew Perroy and Erina Malarkey, Talisa Dever from JVM Realty described exactly that: zooming from the refrigerator in one apartment out to the market around her property, and back again.
So this map follows one slow fridge repair up to the portfolio. It starts on twelve apartment properties around Dallas, each with its renewal rate and how long its repairs take. In the slowest one, the fridge in unit 311 needed three tickets in six months, and the last fix took nine days. That resident gives notice. Seventeen apartments in the building have the same slow fixes, and the property ends up with the lowest renewals in the portfolio, 41% against 55%.
It works the other way too. A resident whose dishwasher gets fixed the same afternoon renews, and when they move to another city where the same operator has buildings, they look there first and bring a friend along. That revenue never shows up next to a work order unless both sit on the same map.
Claude did most of the legwork, and nearly all of it through the @smplrspace platform: generating twelve buildings in the style of real US multifamily, furnishing the apartment, loading the operating data and drawing it on the map and inside the building. It's a good test of our work on making the platform agent friendly. Claude also cut the podcast clips with captions and animated it all. I project managed, directed the data generation and did the art direction, and I did most of it from my phone 🤯
The buildings and their data are synthetic. The places, offices and apartment complexes around the hero building come from OpenStreetMap.
1
118
The episode behind the clips, with Talisa Dever of JVM Realty and Erina Malarkey: Unlocking value with data visualization in multifamily real estate, Spatial sessions Ep 1. youtube.com/watch?v=sHdepLA-…
20
Day 12 of the #30DayMapChallenge is Power. In the last year, power has mostly meant AI data centers. The map opens on the data centers the world built up from 2000 to 2020, then the big AI campuses switch on year by year, and what is being built and announced takes it to 2030: about 3,100 sites from open data, sized by the power they use.
31 of the big AI campuses came online in the last twelve months, 26 of them in the US. Almost 600 more data centers are being built or announced, and outside the US, India and Southeast Asia have the most.
The International Energy Agency expects data centers to use about twice as much electricity in 2030 as in 2025, with China growing the fastest. China was also the hardest country to map, because almost no Chinese data center says how much power it uses, so I worked from the government’s own count of server racks.
Claude did most of the legwork: finding the open datasets, placing about 600 planned sites on the map, and building the animation. I chose the question, pushed for real global coverage when the first sources left China nearly blank, asked for an estimated location whenever a site’s exact one wasn’t clear cut, and did the art direction. As always, it’s built on the @smplrspace map viewer.
Data: Megawattlas, Epoch AI and the International Energy Agency (CC BY 4.0), plus MIIT, CAICT, Bitkom, CSO Ireland and CBS.
1
95
Day 11 of the #30DayMapChallenge is Senses: Sound. I was at a loss for ideas, and then I remembered @Akenza_IoT's post from last year, so I decided to feature it.
At IoT Sparks in Valencia, akenza turned the venue into a live map of sound. @MilesightIoT sensors streamed decibel readings over LoRaWAN, and akenza's dashboard drew them on a 3D floor plan of the building, built on @smplrspace.
The peaks followed the program: presentations, lively discussions, networking sessions. Noise works as a rough map of where people are engaged.
Read more about it on their blog: akenza.io/blog/iotsparks2025.
#30DayMapChallenge #NoiseYouCanSee #IoT #dataviz
2
49
Day 10 of the #30DayMapChallenge is Prompting only, so there is no code and no scripted typing in this one: a real chat, asking a mall questions about its own data, with every answer drawn on its floor plan.
On Day 8 I showed a mock-up of this and said we'd make it real within a year. Here's a first look, three days in. It's an upcoming product the team is building on the @smplrspace platform, connected to AI agents over MCP (Model Context Protocol, the open standard for plugging AI assistants into other tools). It isn't even in beta yet. It works, with a lot of duct tape, and we're still productizing it. I'm kinda leaking this 😱.
I asked the questions an owner asks every month. Which shops pay too much rent for what they sell? Who pays a lot of rent for little footfall, and who pays little for a lot of it? Which leases end in the next six months, and which of them sit in the best spots?
The Shoppes is a demo mall. Its tenants, rents and footfall are synthetic, laid over the real floor plan of mall in Singapore.
The clip has three answers cut from a 13-minute session, 8 questions total. Happy to share the full recording if you want to see the rest, ping me.
#30DayMapChallenge #PropTech #RetailRealEstate #SpatialIntelligence #AIForRealEstate
1
57
Day 9 of the #30DayMapChallenge is Urban-Rural. Singapore is officially all urban, so I went looking for the nearest thing to the countryside I grew up in.
I grew up in rural France, and what I miss most is the pace, the space, and the peace. This map scores every spot on the island where someone lives for how peaceful it is, on hexagons about 110 meters wide.
The score leans most on how densely people live, so a street of houses ranks above a block of apartments. Then come distance from expressways, how much green is around, and a little for low buzz from shops and food. Color and height both follow the score: the busy places stay low and deep blue, and the peaceful ones stand up pale green. Where no one lives, nothing is drawn.
The six places marked are where the very top scores cluster: Shamrock Park, Mount Pleasant, Sembawang, Nassim Road, Adam Park and Canterbury Road. They are streets of houses with space around them, and the farms of Lim Chu Kang barely show up, since so few people live there.
Built in the @smplrspace map viewer, from @openstreetmap data only. The score is a proxy I designed for the picture, not a measure of quality of life. Claude built the page and the scoring. I chose the question and what goes into the score, and I art directed the project.
#30DayMapChallenge #Singapore #OpenStreetMap #dataviz
1
2
188
Day 8 of the #30DayMapChallenge is Utopia, so I explored what Utopia looks like for some of the companies we help at @smplrspace.
The problem for the retail owners we talk to isn't vacancy, it's having the wrong tenants: a premium brand in a value zone, a fourth cafe in the same corner, a shop that lives on passing trade in a dead end. My idea of utopia for them is asking questions in english to your portfolio and getting answers on the map, with the reason and a suggested fix, instead of waiting for a report.
The video starts on a portfolio of retail buildings, drops into the weakest one, and ends with five swaps that take the floor from 78% to 95% of what it could earn.
A big part of what we're building right now is letting users map their data to our floor plans, then plug their own AI with its context, so they can ask questions like this one. This is a mock-up of that. The typing is scripted and the answers are computed from the data. The tenants, the sales and the scoring rule are all made up. The floor is real: a mall in Gurugram, with its units computed from the walls of the digitized building.
AI is very much part of how this could happen, and of this video: it built the page and the scoring. I picked the question, the building and what the answers say.
It's Diwali today, and this one happens to be set in India. Happy Diwali to everyone celebrating.
#30DayMapChallenge #PropTech #RetailRealEstate #SpatialIntelligence
1
75
Day 7 of the #30DayMapChallenge is the 10-minute map. Kinda relieved because I'm a day late and so I have 2 to post today 😬.
10 minutes.. Sounds like a challenge? Well, erm. I took a building that was already digitized and put it on the map in not even two minutes.
Many of our users have digitized buildings they use in 2D or 3D, and a lot of them don't know the same model can be georeferenced in the @smplrspace app and sit on the map right away. This one is City Green Court in Prague. The recording is sped up, but the timer is the real one.
AI's only job on this one was finding me a suitable timer app.
#30DayMapChallenge #aintnobodygottimeforthat
1
80
Day 6 of the #30DayMapChallenge is Vintage, so I drew five Midtown Manhattan buildings the way an insurance surveyor would have in 1890, then switched to how the same buildings look today.
From the 1860s, companies like Sanborn drew every building in a city for fire insurers: footprint, use, materials, number of floors. That was building-level property intelligence, sold by subscription and kept current by pasting corrections over the old sheets. The look of this map comes from those sheets, but every line is drawn fresh. The originals used colors for building materials, here they show what the building is used for.
The second half is the same buildings in the @smplrspace map viewer, where nobody pastes anything over anything. Change the data and the map changes with it.
@claudeai wrote the page. It turned a @Mapbox style into a sepia, paper-textured basemap, colored the buildings by use with Smplrspace's own shell layer, a feature built into the map viewer, and set the hand-lettered labels as persistent tooltips. I chose the buildings and the colors, and the footprints are real Sixth Avenue building shells digitized by our team. The uses and the numbering are made up for the map.
#30DayMapChallenge #CommercialRealEstate #PropTech #Cartography
2
124
There is this line I liked in WeCrashed: "in real estate, it's not about what you can see, it's about who can see you".
That came up a few times with mall owners and managers we work with: can the 3D floor plans show from where a specific shopfront is visible? Day 5 of #30DayMapChallenge is "Sight", so I just built it.
This is one shopfront on the ground level of a mall in Singapore, 8 meters of frontage facing the atrium. Across about 40,000 sqft of floor on two levels, each cell is colored by how much of the shopfront it sees from eye height. 41% of them see at least part of it, the gray ones see none.
I added a line-of-sight check to the @smplrspace viewer for this. In the video, it casts about 300,000 sight lines against the walls, floors and furniture, in about 4 seconds. This function will be released in the next version 🤫.
#30DayMapChallenge #CommercialRealEstate #PropTech
44
I'm joining #30DayMapChallenge a few days late, so days 1 to 4 go out as one map: a retail and office portfolio in India, followed from one city down to a single floor and back out to the whole country, answering the same question at every zoom. Which leases end in the next 12 months?
- Day 1, Points: each building is a point, colored by how much of its area is vacant (red), expiring within a year (amber) or secure beyond that (green).
- Day 2, Lines: the main roads and metro lines that bring people to those buildings.
- Day 3, Polygons: one building, one floor, every unit colored by the same three statuses.
- Day 4, Clusters: the whole country, each city a ring showing how its leased area splits across the three.
Floor plans already carry the geometry most portfolio tools are missing, so one map runs from the country to a single unit. Built in the @smplrspace map viewer using AI, the assets and leases are synthetic. The building and its unit outlines are real, and so are the roads and metro lines, from @openstreetmap.
More real estate questions through November. If there's one you'd want to see on a map, tell me.
#30DayMapChallenge #CommercialRealEstate #PropTech
3
163
Thibaut Tiberghien retweeted
My god this is such a good speech that every SWE needs to hear. You know what? Every person should hear it
Keep the happy memories, eyes on the reality, be excited about the future. That’s the best that anyone can do
See you in Mumbai next week? 🇮🇳
🏬 Own or operate retail malls?
Meet @wanlyntib (CEO) and @tibotiber (CTO) at @MAPIC_India in Mumbai, September 29 - 30, and see how mall teams use interactive floor plans to:
👉 plan tenant mix
👉 track store performance
👉 pitch new concepts
#CRE #Retail
40
Ooooohh, the new observable is pretty hot ❤️🔥
Short compelling notebook by @paulchaikin: when @CP3 joined a team, their win percentage went up. observablehq.com/@pamacha/wh…
49
I've been curious whether Fable 5 is worth the cost, so I gave it something really hard.
We had a problem in the @smplrspace 3D viewer that had been sitting in the backlog for four years. On a really complex building, switching floors meant rebuilding the entire scene from scratch, over 20 seconds of waiting every time. We knew how to fix it, but the rewrite was massive and never worth the disruption given the few spaces this impacted.
We shipped it in two weeks with Fable 5 running almost all of it, 106 commits across 15 pull requests. Floor switches on that same building went from 21.6 seconds to 44 milliseconds.
I wrote up the full story, how it works, what failed the first time, and where I think frontier models actually earn their cost.
t15n.io/the-rewrite-we-avoid…
2
75
We've seen a meaningful shift in the size and complexity of what customers are running on Smplrspace. With this update, the platform is catching up. Proud of what the team has shipped here 🚀
Smplrspace now handles enterprise operational complexity:
👉 One billing account across organizations
👉 Use case agnostic project scoping
👉 Usage analytics at every level.
➕ A stack of editor and SDK upgrades
Built for teams running more than one of everything.
#proptech
1
1
37




