Landing page
Item categories
Graph visualiser
Analysis
Highlighted connections
Inventory

SubGraph

Inspiration

Shopping platforms compare products by brand, features, reviews, and price, but they rarely answer a more personal question:

What does this add beyond what I already own?

Products that look different can perform many of the same functions. An air fryer and a countertop oven are separate products, for example, but several of their capabilities may overlap.

SubGraph identifies this functional overlap. It shows what is already covered, what is genuinely new, and what the user is paying for the difference.

It does not tell people whether they should buy something. It gives them evidence to make a more informed decision.

What it does

SubGraph turns a household inventory into an interactive capability graph.

When a user evaluates a product, SubGraph:

Breaks it into three to eight functional capabilities.
Compares those capabilities with every item they already own.
Calculates how much of the product is already covered.
Identifies the capabilities that are genuinely new.
Shows the cost of each new capability.
Maps every result to an inspectable edge in the graph.
Suggests a cheaper or secondhand way to obtain the missing capability when appropriate.

For example, a countertop oven may overlap heavily with appliances already in the kitchen while still adding one useful function. SubGraph shows both the overlap and the new value instead of returning a simple buy-or-skip verdict.

Signed-in users can photograph a room, review the detected items, and save a personal inventory. Detected items remain provisional until confirmed. Images and temporary scan data are discarded after review; only confirmed inventory fields are stored.

Three built-in demonstrations work without signing in:

A countertop oven with substantial overlap.
A completely redundant fourth USB-C cable.
A mini drone with genuinely new capabilities.

How we built it

Technology	Purpose
React and TypeScript	User interface and application logic
D3 and SVG	Interactive force-directed capability graph
OpenAI structured outputs	Product and capability decomposition
OpenAI vision	Provisional item detection from room photos
OpenAI embeddings	Semantic capability matching
Clerk	Authentication
Neon Postgres	User inventory storage
Neon pgvector	Capability embedding storage and lookup
Drizzle ORM	Typed database access and migrations
Vercel Functions	Server-side evaluation, photo processing, and inventory APIs
Vercel	Hosting and deployment

SubGraph represents capabilities using consistent verb-object phrases such as:

toasts bread
crisps food with hot air
charges usb-c devices

It first checks for exact matches. Unmatched capabilities are embedded and compared with the active inventory vocabulary using cosine similarity.

Coverage is directional and weighted:

$$ \text{coverage} = \frac{\sum_{\text{covered}} w_{\text{specificity}} \cdot w_{\text{tier}}} {\sum_{\text{all}} w_{\text{specificity}} \cdot w_{\text{tier}}} $$

Primary capabilities receive more weight than secondary ones, while capabilities shared by many products receive lower specificity weights. This prevents generic functions from dominating the result.

The score is directional because it measures how much of the proposed product is already covered by the user’s inventory, rather than how similar two products are.

D3’s force simulation creates clusters from shared capabilities instead of fixed positions. The graph and verdict list are generated from the same data, so every result can be traced to a visible connection.

Challenges we faced

Defining functional overlap

Product-name matching was too shallow, while unrestricted AI output produced inconsistent capability descriptions.

We combined structured model output, a strict capability naming convention, exact matching, embeddings, and an inventory-specific vocabulary.

Preventing misleading matches

Generic capabilities could make unrelated products appear more redundant than they actually were.

We addressed this with specificity weighting, primary and secondary capability tiers, and directional scoring.

Making the graph useful

The graph needed to explain the result rather than act as decoration.

Every verdict row and graph edge is derived from the same underlying data. Users can inspect which existing item covers each capability.

Handling photo recognition responsibly

Room scans may contain incorrect detections or private information.

Detected items remain editable and are never added automatically. Images, confidence values, model evidence, and temporary review data are not retained.

Keeping the demo reliable

Live model and database requests can fail during a demonstration.

We added deterministic offline examples that use the same scoring and graph logic as live evaluations.

What we learned

We learned that directional functional coverage is more useful than generic product similarity.

The important question is not whether two products resemble each other. It is how much of the proposed product is already covered by the user’s entire inventory.

We also learned that visualization can act as an audit trail. Mapping every result to a graph edge makes AI-assisted recommendations easier to understand and verify.

Most importantly, software designed to reduce unnecessary consumption must still be able to show when a purchase adds genuine value. SubGraph supports better decisions rather than blocking purchases.