02 / Civic data / NYC Parks
Make the data useful
where decisions happen.
A data product begins with a question in the world. The work is to connect that question to information, assumptions, and an output people can use.
Where could
a tree be planted?
At NYC Open Data Week on March 25, 2025, Shawn presented Finding Green Spaces: A Model for Tree Planting in NYC as a Data & Product Designer with NYC Parks. The event describes a model bringing together Tree & Site data, planimetrics, traffic signs, and the Heat Vulnerability Index to support site selection.
The product problem
A location is more
than a coordinate.
Planting opportunity combines the physical setting, existing assets, constraints, and the reason to prioritize one place over another.
The design challenge is to bring different kinds of information into one decision process while keeping their meanings legible. A geometric feature, an operational record, and a vulnerability measure answer different questions. Combining them requires judgment about the decision each input is allowed to influence.
This is an explanatory reading of the model’s input categories, not a reproduction of its rules or an interactive planting recommendation.
A working scene / 03
Read the block in layers.
What changes when different kinds of information meet the same street?
01 / Geometry
Begin with the physical setting
Planimetrics describe the shapes and edges of a place. The first sheet asks what space is physically present.
02 / Existing conditions
See what is already there
Tree & Site data bring existing assets and site records into view. They answer a different question from street geometry.
03 / Constraints
Check what could limit use
Traffic signs are one named input to the public model. This sheet stands for operational checks, without stating any siting rule.
04 / Priority
Ask where it could matter most
The Heat Vulnerability Index adds context for prioritization. The lens poses a question; it does not show a score, threshold, or recommendation.
A visual comparison / authored interpretation
Different material.
The same design question.
What must be brought together, what can the system produce, and who decides what to do with it? Select a stage to compare the two settings.
01 / Question
Where might a tree be planted?
The public presentation describes a site selection model meant to help NYC Parks foresters find promising planting locations. The question comes from a real operating need.
Limit / A model can organize the search; it cannot make a planting decision.
What work is worth doing?
A Human direction becomes useful when it is framed as bounded work with a result that can be judged. Corpus keeps an accepted requirement distinct from the result.
Limit / A request is a contract for work, not proof that the work succeeded.
02 / Ground
Bring unlike data into view.
The event names NYC Parks Tree & Site, NYC Planimetrics, DOT Traffic Signs, and the DOH Heat Vulnerability Index as model inputs. Each layer describes a different part of the planting question.
Limit / The event listing does not publish the model’s rules or the quality of each input.
Read the contract and the source.
Agents work from an accepted scope, the code or material it touches, and evidence gathered during investigation. Those objects carry different claims and should stay traceable.
Limit / Stored evidence is a starting point for judgment, not automatic truth.
03 / Make
Produce candidate sites.
The described model combines city data to identify promising tree planting sites. This drawing expresses that product purpose; it is not a reproduction of its spatial method or a real map.
Limit / A candidate location is an output for consideration, not a planting recommendation from this page.
Produce an inspectable candidate.
An agent investigates and implements a bounded change. Source, tests, and a handoff give a reviewer a specific candidate to inspect.
Limit / The maker’s completion report cannot grant independent acceptance.
04 / Judge
Use professional judgment.
The model’s stated purpose is to support foresters’ site selection. On site conditions and planting decisions remain with the people and process using its output.
Limit / The public event record does not establish which sites were planted or any measured canopy effect.
Review, then record the result.
An independent reviewer judges the delivered candidate against the accepted work. The outcome and its evidence can then be carried forward for another reader.
Limit / Review establishes a bounded judgment about the candidate; it does not make every broader claim true.
Ground: NYC Open Data Week’s tree planting presentation · How Corpus work moves · Judging the work. This is an authored comparison, not a depiction of actual sites, model rules, or a live agent execution.
A collaborative contribution
Help shared data
hold up in another use.
The Nature Conservancy’s 2026 practical canopy release, published August 26, lists Michael L. Treglia and Karen Leu as authors. It acknowledges Shawn, Uma Bhandaram, and Tyler Gibson at NYC Parks for thought partnership and support in the robust use of Parks datasets and potential street tree locations.
That is a contribution through collaboration and domain knowledge. The dataset authorship belongs to the named creators.
Read the release and acknowledgmentsA dataset becomes more useful when someone understands the operational world behind its fields. That knowledge matters when the data moves into a different model or decision.
What carries into AI / Portfolio perspective
Understand the boundary.
Then build the system.
The connection to Corpus is a design practice: translate a real problem into a system with inspectable inputs, responsibilities, and outcomes.
Start with the decision.
For civic data, ask what a planner or forester needs to decide. For agent work, ask what a maker, reviewer, or successor needs to know. A useful system organizes information around that decision.
Keep meanings distinct.
Different data layers carry different information. In Corpus, a source, a claim, a contract, a receipt, and an accepted result also need distinct treatment. Combining them should preserve what each one can establish.
Design for use and revision.
Outputs meet people and operating conditions. A workflow must remain understandable when assumptions change, another person takes over, or a tool behaves differently than expected.
The broader direction
Bring domain judgment
into AI development.
This portfolio brings together a public-sector data practice and an ongoing agent systems experiment. The value of that combination is attention to the whole working situation: the problem, the information, the user, and the mechanism that connects them.
Corpus extends that attention to a new setting, where an agent can produce work but the surrounding system must help people judge it and continue it.
Read the Corpus case