Inspiration

Have you ever made an impulse purchase, like buying a top you wore only three times before tossing it? We have, and in the moment it felt harmless. But every second, the equivalent of one garbage truck of textiles is landfilled or burned, and every piece in that pile took real water, energy, and labor to make. Greenpeace Germany tested 56 Shein garments, and 18 (32%) exceeded EU REACH chemical limits (Greenpeace), so some of what we wear against our skin isn't even safe.

We wanted a friend's second look right when you tap "add to cart," because one more moment of thought is the cheapest thing we can give the planet, our bodies, and our wallets.

What it does

  • A Chrome extension reads the product page and shows a card with one of three verdicts: have this look, micro trend, or good buy.
  • It checks your closet with embeddings, so it can say "you already own something like this."
  • It reads the materials, compares them with a materials database, and flags any that could be toxic in contact with skin.
  • If a photo has no stated materials, FashionCLIP gives one hedged estimate, labeled low confidence.
  • The Social Trend Radar matches the item to one of 42 clothing styles measured on Bluesky and shows whether that style is a micro trend, established, or still active.
  • A Photon iMessage agent gives the same second opinion from a photo, in a friend-style reply.

How we built it

  • Extension and backend: a Chrome MV3 extension (Shadow DOM card) and a FastAPI backend.
  • Data and AI: MongoDB Atlas with vector search. Claude writes the verdict wording. Vultr Serverless Inference (qwen3-embedding-4b) embeds the Bluesky posts. Gemini embeds users' closet items, understands items in real time, and acts as a fallback when scripts fail to run.
  • Radar: a graph of 25,658 authors, 226,251 edges, 502 communities, and 42 clothing styles. To decide whether a style is a micro trend, established, or still active, we developed two methodologies. A Deep Agent decides each style's outcome with tools, skills, and a per-style tool budget. A separate rule baseline is kept for comparison, and we analyzed where the two disagree.
  • Photon: an iMessage agent.

Challenges we ran into

Chrome integration. Building a Manifest V3 extension that works inside a real shopping page was harder than we expected. Each store lays out its page differently, add-to-bag buttons load late or change after you click, and the extension has to stay out of the way while it reads the product, checks your closet, and shows a verdict. Debugging inside someone else’s page, with strict extension rules on what code can run and where, took a lot of trial and error.

Finding insights that are not obvious. Collecting social data was the easy part. The hard part was getting something out of it that you couldn’t already see by scrolling the feed. A chart saying “this style is popular” tells nobody anything new. We spent a lot of time on what the data could say that a person could not eyeball, such as which styles are saturated, which are quietly fading, and where our own evidence is too thin to call, so the radar says “unknown, low confidence” instead of guessing.

Making sense of every brand’s product data. Hollister, PacSun, and every other brand describe their products differently, with different names, categories, material labels, and page structures. Parsing all of that into one consistent format, so the same tee from two brands can be compared against your closet and the trend data, was slow and messy work.

Accomplishments that we're proud of

While figuring out which categories different fashion styles belong to, we analyzed every case where the agent and the rule baseline disagreed. Because we gave the Deep Agent sufficient tools, it pulled more from the Bluesky posts than the rules could: materials, the general consensus on a trend, and how established the creators who started it were.

The disagreement analysis surfaced a lot of unique insight. One example was balloon pants. Their popularity has been relatively stagnant, yet the agent understood they are still considered fast fashion because of the materials most companies use to make them. We found many results like this one in our social media analysis.

What we learned

We learned a lot about the tools we built with this weekend, and most of them were new to us:

Nessie by Capital One, for working with banking data and savings goals LangChain’s deepagents, for an agent that reasons through each style’s evidence and decides its trend outcome Arlequin AI’s approach to social media analysis, which combines content with connection patterns Amazon’s tools for building practical AI Google Gemini, for understanding items and pages and for embeddings Vultr, for hosting and running our data work MongoDB Atlas, for storing and querying our data Photon, for putting Second Look inside iMessage Chrome extensions, which taught us the most about working inside someone else’s web page

Along the way, we learned how to pick up an unfamiliar product quickly: read the docs, build a small working piece first, and connect it to everything else before polishing. We now feel comfortable picking up new tools whenever a project needs them.

What's next for Second Look

More data sources beyond Bluesky. We want to pull in more places where people talk about fashion, then let the deepagent weigh them against each other. If a style looks hot on one platform and dead on another, the verdict should say so.

Learning from your feedback. Every time someone likes or dismisses one of our suggestions, that is a signal. We want to feed those reactions back into the system so the closet matching, the trend verdicts, and the “skip it” nudges get sharper for each person over time. A “that’s not the same one” or “good call” tap should make the next second look better.

Recommendations that match your style. Second Look should learn what you actually like. If you keep saving minimalist neutrals or loving a certain vintage look, we want it to suggest more of that and build outfits around the pieces you already own. The goal is to steer you toward things you will wear, so you buy less and waste less.

All three aim at the same thing: fewer impulse buys, less waste, and a closet you actually wear.

Built with

Chrome extension, FastAPI, MongoDB Atlas, Claude, Gemini, Vultr Serverless Inference, Photon, FashionCLIP, NetworkX, Bluesky

Sponsor Tracks

Best Usage of Nessie (Capital One)

Second Look uses Nessie to make "Skip it" move real money. When a shopper skips an item on the add-to-bag card, our backend transfers the item's price from their Nessie checking account to a Nessie savings account tied to a goal they set at signup (in our demo, $2,000 for a Japan trip). The card first shows "Skipped. Savings update pending," then the popup updates its "money saved by skipping" total (the demo shows $55.67 of $2,000) and the goal progress bar.

We use Nessie accounts and transfer endpoints, so each skip is an actual sandbox transaction with an audit trail. We log the item and price so the savings total can be traced to specific decisions.

Track 2: Social Media Analysis, Combining Content & Connectivity Patterns (Arlequin AI)

We built a Social Trend Radar on Bluesky that uses both layers to judge how long a clothing style lasts.

  • Content: we embedded posts with Vultr's qwen3-embedding-4b and grouped them into 42 styles.
  • Connectivity: a NetworkX graph of 25,658 authors and 226,251 edges (10,398 interaction edges, 215,853 embedding-similarity edges) with 502 communities.

Both layers mattered. Interaction edges alone left the largest connected component at 2.2% of authors, while adding similarity edges connected 97.3%.

Per style we measure weekly share over 53 weekly windows, plus how fast it spreads (weeks to reach 3 communities) and how concentrated it is in a few communities, used only when that data exists. Outputs are explorable through a graph view and a per-style timeline. We were able to generate unique insights about different fashion trends and how they travel across a network.

Limits: the data is a sampled lower bound with a 20-post weekly cap, and 15 styles saturated the cap in more than 25% of weeks, so we flag censored peaks.

Results: recall@1 = 0.87, recall@3 = 0.95, recall@5 = 1.00

Best Use of Deepagents (LangChain)

A deep agent decides each style's outcome (micro trend, established, still active, or unknown), and that label is the only one the app shows. We built it with create_deep_agent, a filesystem backend, and four skills (style-judge, precision-auditor, disagreement-analyst, analyst-note-writer), plus measurement tools in scripts/radar_agent_tools.py. Each style gets a budget of 12 tool calls.

The agent writes decisions through a guarded write_decision tool that enforces three rails:

  1. Enough data: under 100 posts or under 8 weeks of data means only "unknown."
  2. Evidence IDs required.
  3. Lifespans only from tool-measured runs with a sourced URL.

Decisions persist in MongoDB as they happen, so a run can resume. We also kept a rule-based baseline and asked the agent to explain every disagreement with it. 34 of 42 styles agreed, and we explained one example disagreement (balloon pants) above.

Best Practical AI Innovation (Amazon)

Second Look is a second opinion at the moment of purchase, for shoppers who buy things they already own or trends that die in weeks. A Chrome extension reads the product when you hit add to bag, and the agent takes several steps:

  1. Matches the item against your photographed closet (vector search).
  2. Reads the real fabric from the label and estimates wear and cost per wear.
  3. Builds three outfits from what you own, or says the piece is not versatile.
  4. Checks the style against the trend Radar.
  5. If you skip, moves the money to savings through Nessie.

It adapts to the user through their closet and savings goal, and reaches shoppers in stores through a Photon iMessage agent. Target user: students and young shoppers who overbuy.

Build a Messaging Agent using Photon 2.0 (Photon)

Second Look reaches shoppers who are in a store, with no browser open, through an iMessage agent built on Photon. You text a photo of something you're thinking about buying, often with a question like "What do you think of this top?", and the agent replies in the thread like a friend would, with a short, casual second opinion instead of a report.

It flags fabric and chemical-safety concerns to check on the tag using a materials identification model, and it points out when the piece looks like something already in your closet, using the same closet matching as our Chrome extension.

The messaging agent is the second front door to one shared backend, so a skip on your laptop and a text from the fitting room draw on the same closet and the same trend Radar. Photon let us ship this over iMessage, where our users already are, without building an app.

[MLH] Best Use of Gemini API

Gemini is how Second Look understands what you're about to buy and what you already own. Claude writes the final verdict wording, and Gemini does the reading and matching underneath it.

  • Item understanding: Gemini turns a shop item's photo and text into structured attributes (category, color, fabric, cut, and style). When a store photo shows a full outfit, it draws a box around the target item so we compare the right garment.
  • Page reading: when a store page is hard to parse, Gemini pulls the main product out of the page's trimmed text.
  • Closet intake: Gemini extracts garments from the photos and video frames a user uploads (10 images per call), then merges duplicates in a text-only follow-up call.
  • Embeddings: gemini-embedding-2 at 768 dimensions embeds each item and closet piece. A new item's embedding is matched against your closet through MongoDB Atlas Vector Search, which is how "you already own this look" works.
  • Trend agent and evals: our trend-agent script defaults to gemini-3.5-flash on failure, and we used Gemini for attribute and verdict evaluations and for tuning the matching thresholds.

[MLH] Best Use of Vultr

The Radar's content layer runs on Vultr Serverless Inference. We embed the Bluesky posts with qwen3-embedding-4b and VultronRetrieverPrime-Qwen3.5-8B, and those vectors drive the style grouping, the similarity edges in the author graph, and the matching of a shopper's item to one of 42 styles. We did 31,599 image and text embeddings.

We tried many different embedding models on Vultr. It was easier to work with, and running local models kept costs lower. With GPU credits, we would train our own multimodal model specialized for clothing items.

[MLH] Best Use of MongoDB Atlas

Atlas is the system of record. Two Atlas Vector Search indexes do the matching:

  • closet_vector uses $vectorSearch to match a scanned item to the shopper's closet (45 closet items indexed across 21 demo accounts).
  • social_posts_vultr_vector matches it against 31,599 embedded Bluesky posts to find its style.

Separate collections hold:

  • social_trends: measured stats for 42 styles
  • trends: the records the verdict engine reads
  • radar_decisions: each style's rule label, agent label, final label, source, and agreement flag, for 42 styles across 3 runs
  • item_materials: per-item material estimates, upserted per scan

We write results as we go, so a failed run resumes from where it stopped instead of restarting.

Built With

  • bluesky
  • chrome
  • claude
  • deepagents
  • extension
  • fashionclip
  • fastapi
  • gemini
  • gemini-3.5-flash
  • imessage
  • inference
  • langchain-deep-agents
  • mongodb
  • mongodb-atlas
  • mongodb-atlas-vector-search
  • nessie
  • networkx
  • photon
  • qwen3-embedding-4b
  • serverless
  • shadow
  • vultr
  • vultr-serverless-inference
Share this project:

Updates

Submission history