Phia Sniper

## About the Project
(For a more technical breakdown, please refer to our readme.md file)

### Inspiration

To win at online resale fashion, you have to scroll — constantly, 
repeatedly, for the slim chance of finding that diamond in the rough. 
Resale moves at the speed of social media, but the tools haven't caught 
up. Saved searches notify you too late, offer no way to incorporate user 
feedback, and completely fall apart when you don't know exactly what 
you're looking for yet. Sometimes you don't want a specific item — you 
want a feeling. A Pinterest board full of Mediterranean summer fits. A 
Mamma Mia vibe for a Greece trip. No saved search can handle that. We 
built Sniper to close both gaps.

### What We Built

Sniper is an AI agent layer built on top of Phia that shops resale 
continuously on behalf of users. The core insight is that shopping intent 
exists on a spectrum — from "I know exactly what I want" to "I have a 
vibe and need help getting there" — and no existing tool serves the full 
range.

We built two agent types to anchor each end of that spectrum:

**High-Intent:** The user specifies exact criteria — brand, 
item, size, condition, price ceiling — and the agent loops continuously 
until it finds a match. Not a saved search. A saved search waits. This hunts.

**Low-Intent:** The user describes an aesthetic in natural 
language or pastes a Pinterest board URL. The agent interprets the vibe 
and surfaces resale items that match it, ranked semantically rather than 
by exact criteria.

Users can run multiple agents simultaneously, each with its own toggle, 
configuration, and live matches feed.

### How We Built It

We started with the matching problem. For the high-intent agent, the 
challenge is speed and precision — polling listings frequently enough 
to matter in a market where the window on a coveted item can be measured 
in minutes. For the vibe agent, the challenge is semantic: translating 
something as fuzzy as "Mamma Mia energy" into a ranked list of real 
resale listings requires genuine aesthetic reasoning from the model, not 
just keyword matching.

We used an LLM to handle the vibe translation layer — converting natural 
language descriptions and Pinterest board content into search parameters 
and ranking criteria. The agent loop runs on a polling architecture, 
continuously querying listings and running new results through the 
matching logic while the toggle is active. The multi-agent dashboard lets 
users manage their full portfolio of running agents in one place.

### Challenges

The hardest problem was the pivot. We originally built a single 
low-intent vibe agent. After presenting to Sophia and Phoebe, the 
feedback pushed us toward something more high-intent — an elevated saved 
search that runs continuously and delivers real-time notifications when 
an item is sniped. Going back to the drawing board, we realized the 
stronger product wasn't one or the other — it was both. Users should be 
able to run multiple agents at once, each directed at a different point 
on the intent spectrum. Rebuilding the dashboard architecture to support 
this mid-hackathon required rethinking core assumptions fast.

The second hard problem was vibe matching. Exact matching is 
deterministic — a size large either matches or it doesn't. Vibe matching 
is probabilistic, and getting the model to rank results in a way that 
feels genuinely aligned with a user's aesthetic — rather than just 
topically related — required careful prompt engineering and iteration on 
how we represented intent to the model. This also surfaced the need for 
user feedback loops: on Felix's suggestion, we added a feedback field so 
users can refine what the agent knows about their taste over time.

We also had to think carefully about the agent lifecycle — what it means 
for an agent to be "running," how frequently it should poll without 
hammering APIs, and how to surface new matches in a way that feels live 
rather than static. For the purposes of this demo, the agent runs to 
completion upon toggling on. In a production environment, this would be 
backed by cluster-managed deployment jobs running chronically in the 
background.

### What We Learned

Intent is a powerful frame for thinking about shopping tooling — and most 
existing products do not give it enough attention (we almost didn't either!). A shopper hunting a specific grail and a shopper building a vacation wardrobe need fundamentally different 
tools, and the gap between what they need and what exists is large enough 
to build in.

We also learned that great AI products aren't just about the model — 
they're about the partnership between probabilistic and deterministic 
tooling. The LLM handles the fuzzy, aesthetic, judgment-heavy parts, such as receiving a Pinterest board turning it into a powerful prompt. 
Deterministic systems handle the precise, reliable, structured parts, such as scoring matches numerically. 
Getting that balance right is what makes the experience feel both 
intelligent and trustworthy.

Built With

• kernel.js
• next.js
• playwright
• supabase