fashion tech directory

I kept running into the same problem: I'd hear about a new fashion tech company, go to look it up, and it either didn't have a LinkedIn page yet or it was basically invisible on search. And even when I did find it, "are they hiring?" turned into a manual scavenger hunt across random job pages. So I built a small thing I genuinely needed: a directory that makes discovery + filtering actually fast.

the problem (in plain English)

Fashion tech is early. A lot of companies aren't indexed well, don't show up cleanly on LinkedIn, and don't fit neat categories. I wanted one place where I could answer: remote-friendly AI styling startups that are actually hiring right now, without 10 tabs open.

what i built

Fashion Tech Directory is a web app with 78 companies right now. Each company has structured fields (category, business model, stage, location, hiring status) plus a short description so results aren't just a dead list. Filters narrow the set, then search ranks what's left.

Demo GitHub

search (keyword + semantic)

I treated this like a retrieval problem and shipped two modes:

Keyword search: a fast baseline when the user has exact intent.
Semantic search: embeddings + cosine similarity so phrasing doesn't have to match perfectly (ex: "eco friendly" can still surface "sustainable materials" companies).

semantic search runs in the browser

This was important to me: no external API calls. I run all-MiniLM-L6-v2 locally in the browser using @xenova/transformers. Company embeddings are precomputed at build time, then the user's query gets embedded at runtime and ranked by cosine similarity.

annoying edge cases (aka what made it real)

Careers link 404s: not every company uses /careers or /jobs, so I stopped guessing and made careers links an explicit field, with guardrails so the CTA falls back gracefully when it's missing.

Exact-match failures: keyword search alone wasn't enough, which is why semantic retrieval became a real upgrade.

how i checked myself

I kept an evaluation folder and a small query suite so I could re-run searches and make sure changes actually improved top results instead of just "feeling better."

next steps

If I keep building: automated link validation, a bigger evaluation set, and lightweight tests around filter/search correctness as the dataset grows.