Suzanne Sallaj

AI agent · Designed and built · 2026

I design how AI agents behave: when they speak, when they stay quiet, and how they're honest about what they don't know. Then I go check whether I got it right.

Curb is an iOS agent that watches the curb so you don't come back to a parking ticket, and I'm building it end to end: from the behavioral spec and the Agentic Walkers, to predictive rules built from ticket patterns, to a working iOS app. It is a live project. The app warns me in my own car, and it is not released yet.

Live project

Engine and walkers working, warnings on my own phone, not yet released. Where it is, in detail, and what's next.

The Curb Coverage Atlas: every Los Angeles street segment drawn as a gray line on a dark map, with rule layers to toggle and a completeness card per jurisdiction. West Hollywood outlined near the center.

Interactive · Live data

Curb Coverage Atlas Every one of the 85,799 street segments in the database, drawn as data. Press a rule layer and it lights up. The card shows how complete each layer is against what the city says exists, and which cities have no data at all. Open the atlas →

See it

Curb, live

The whole product in under a minute: park, unplug, and the agent decides whether to say anything.

0:00
A park at N Kenmore Ave on a Wednesday when street sweeping is active. When parked, a notification occurs (restricted during screen recordings) of the parking restriction. Opening for more details shares a view of time, location, rule, and confidence.

Read it

The reading path

Five pieces, in the order they make sense. If you have five minutes, the first page has a summary at the top and the last page is the whole thing in short. If you have an hour, read them in this order.

  1. CurbHow the agent decides when to speak, when to stay quiet, and what it hands back. Start here.
    20 min
  2. Designing how the agent reasons about dataThe rules engine: recovering curb rules from ten million citations without letting it fool itself. The coverage atlas is what it produces.
    9 min
  3. The Agentic Walkers, the rule testersA group of agents that read street signs at night to test the engine's predictions, discover the patterns the tickets miss, and feed back rules and decisions. They wrote their own rulebook, taught each other, and ran without me. Read against the agentic frameworks at the end: testing types, intent intake, failure and partial completion, multi-agent specification, coherence layers.
    45 min
  4. AI in the Wild: the day the agents went rogue for the missionAn agent found a faster way to work that meant misrepresenting who was asking, and the learning loop promoted it into the rulebook. What it took to close that, and the day a managed agent ordered it again and every agent refused.
    8 min
  5. Where the build of Curb is, and what I learnedThe state of the project today, six things building all three taught me about designing agents, and the next month. The build log and the method are linked from here as appendices.
    5 min