02
Doubland
ai personas with memory, routines, and realtime replay inside a social sim
Doubland is my most important freelance project with a client: eight months so far, working inside their team. It is a product by Doubland, Inc. in Pittsburgh: you create an AI Double based on your personality and watch it live in a town with other Doubles, as a show about you and your friends. I joined a product that was already moving, and my work covers how the Doubles move, arrive and start conversations, the data behind every place in the town, a replay you can trust, performance on phones, and today the front end, the landing of doubland.ai and the live map.
- most important freelance client engagement, eight months
- AI Doubles of real people, live 24/7 with replay
- agent movement, conversations, place data, replay, performance
- owner of the front end, the landing and the live map
Product
Doubland is where the question "What if?" gets answered by your Double, a faithful snapshot of a real person: their personality, voice, decision patterns and social instincts. The Doubles of a group of real people live together in one town for a season; in the current Survival format a nightly vote sends one of them home, and every day becomes an episode. You can watch live 24/7, follow any Double and replay every moment, and a Showrunner can cast their own group in a private run. It is not a game and not a visual simulation of life: the pixel town is only the face, what matters is what the Doubles say and do. It is an invite-only beta for adults.
Role: eight months, from the Doubles to the landing
I have been on Doubland for eight months, inside the client team, and my scope grew the whole time. I started on the replay viewer and moved into the Doubles themselves: how they find their way through the town, how they reach the exact spot they are heading to, when two of them start talking, and the data that describes every place. Then I built the player around it: timeline, speeds, live mode, share links and highlights. Since September I own the front end and the landing of doubland.ai: the new city map and how it runs on phones, the person card on each Double, chat with a Double from the landing, and account, sign-in and onboarding. My way of working there is to measure first, write down the decision and why, and ship in small steps with tests, checked on real phones.
How the Doubles find their way
When I started, the Doubles sometimes cut through walls on a diagonal, froze in a doorway, or jumped when a replay began. The root problem was that the route a Double followed and the obstacles of the map did not always agree, so I made the route search the one place that decides how a Double walks. A Double waits until the obstacles of the map are loaded before it moves, furniture blocks the way unless it is the thing the Double is going to use, and other Doubles are a cost along the route but a hard stop at the destination, so they walk around each other in open rooms and still squeeze through a narrow corridor instead of freezing. I also made the search itself faster, so long walks across town stopped timing out. The result is Doubles that walk through rooms like people do, with the same route in the live view and in the replay.
Getting to the right spot
A Double that decides to have coffee has to end up at a free seat in the café, not at the door, next to the table or on top of someone else. I reworked how a destination becomes an exact spot: every place knows the spots where its activity happens, a Double settles on one of them and faces what it is using, and when several Doubles head to the same group of seats they are spread over the free ones instead of piling onto one. I also made travel time count: an activity starts when the Double arrives, not when it decided to go, and if a place is too far for the time available the plan makes room for the walk instead of the Double never getting there. Before, Doubles hovered next to their bed or circled a counter; after, they sit, cook and sleep where the action is.