PrecedentAI is a prototype for how designers might steer AI—not just use it.
The bigger challenge isn’t technical, but cultural.
The early draft
Precedent AI bridges the gap between inspiration and creation. Upload a hand-drawn sketch, and the platform—built with Python, Flask, and a custom-trained AI model—searches its database to return visually similar projects. No more scrolling through the same handful of references; instead, discover unexpected connections, forgotten gems, or innovative solutions from a broader design lexicon.
Merging code and craft
This tool is designed for the sketching phase, when ideas are raw but full of potential. By analyzing formal qualities, spatial relationships, and compositional patterns, it surfaces matches that are visually analogous, not just keyword-dependent. Whether you're exploring massing, façade treatments, or plan organizations, Precedent AI helps you break free from creative ruts—not by replacing your intuition, but by expanding your palette of references.
Mission for the future
Built from an architect’s perspective, this isn’t just image recognition—it’s a way to see differently. The goal isn’t replication, but resonance: a tool to help you find inspiration in the unfamiliar, while staying grounded in built precedents.
Turning the page
There's always more work to be done in design and here's what I'm currently working on:
1. Training the AI Architecturally
Beyond Pixels: Incorporate architectural descriptors (massing, thresholds, hierarchy) into similarity metrics. Hybrid Model: Combine ResNet50 with a graph database tagging projects by design concepts (e.g., "courtyard," "cantilever"). Feedback Loops: Let designers "vote" on matches to teach the AI subjective preferences.
2. Expanding the Database
Curated Diversity: Partner with universities to include student work, vernacular architecture, and unrealized proposals.
Temporal Layers: Tag projects by era to trace how certain formal choices evolve (e.g., "brutalism 1950s vs. 2020s").
3. Designing the Interface
Sketch Overlays: Allow side-by-side comparisons with opacity sliders ("How does my sketch align with Louis Kahn’s geometries?").
Concept Maps: Visualize how matched precedents relate to each other thematically.
How it works.
1b
Upload New Data
Upload new data to the resource hub, for Loads Engineers to efficiently and effectively complete their jobs and provide the references for them to onboard with ease.
2
Maneuver Definition
After selecting the loads drop, you will select and/or confirm designated maneuvers to plot the load cases against. Then you will confirm and generate the potato plots.
3
Review and Make Downselections
Once generated, review the automated load case selections and make any necessary updates to potato plot hulls.
4
Export Data
Confirm and final load case selections after reviewing and editing potato plot hulls. Export a CSV and PNGs of data and send to key parties.
1a
Loads Drop Selection
Choose from the list of loads drops in the cloud database or upload a CSV. This will be the load case data that will define potato plot perameters you will downselect from.
Why you ask?
Through Precedent AI, I’ve explored a question larger than the tool itself: What does AI mean for designers?
This project began as a technical experiment—Flask, Python, ResNet50—but became a lens to examine how machine intelligence can augment (not replace) creative intuition.
By training an AI to "see" architectural sketches like a designer would, I’ve confronted fundamental tensions.
Bigger Questions-
Patterns vs. Originality
AI excels at finding formal similarities, but how do we ensure it inspires divergent thinking, not just derivative solutions?
Data Bias
Architectural databases skew toward celebrated works. How might we curate precedents to highlight underrepresented voices or typologies?
The Designer’s Role
Is AI a collaborator, a critic, or a library? (In Precedent AI, it’s a provocateur—surfacing unexpected visual dialogues.)
A New Community
Connecting with ML researchers, architects, and educators revealed shared frustrations—e.g., "Why can’t software understand design intent?"
.svg)
