I built StartupSeeker in my spare time because I kept missing the good stuff. I’d spend hours poking around large databases, only to discover near‑perfect startups buried under brittle filters and exact‑match search. It felt like asking a DJ for tracks that say “love” in the title (keyword) when you mean “soulful songs for a date night” (semantic).
Tell me if this sounds familiar: you spend a night hunting for a small tool, maybe a CLI to compare JSON schemas or a tiny browser helper to copy clean tables. Google feeds you lists and affiliate pages, but not the thing. So you sketch your own, ship a scrappy MVP, feel oddly proud…and then two days later, a friend sends a link to the exact tool you needed, quietly maintained by this startup for a year. It wasn’t missing. It just wasn’t findable with literal keywords. That’s the itch I couldn’t ignore.
And it has to be fast. I’m not sitting through a 30‑minute “AI deep research” spinner to get a PDF‑length answer. Give me meaningful results in seconds, something I can scan, click, and move on with before the coffee cools.
So, I built something that behaves differently. I wanted a search engine that reads meaning rather than just the exact words. The system looks for intent, interprets the ask in plain language, and returns a tidy shortlist that feels like a thoughtful researcher prepared it. Instead of treating you like a form, it behaves like a careful reader, catching context, weighing clues, and choosing examples that actually fit. The plumbing exists, sure, but the real story is simple: you say what you mean, and the results respect it.
You know what? Search should feel simple. Type, get answers. The trick is making the machine do the hard part so the human doesn’t have to.
Semantic Search in Plain English (Embeddings 101)
Here’s the gist. Keyword search matches the letters you typed. Semantic search tries to match the idea behind those letters. It does that with embeddings, compact numeric representations of meaning. Think of each query and each company profile as a point in a huge coordinate system. Points that mean similar things sit near each other. Nearness (cosine similarity/inner‑product) is our signal.
Because those coordinates are just numbers, we can use fast nearest‑neighbor indexes to jump to the closest results, no heavy reasoning loop, no long wait. That’s how you get intent‑aware results in seconds rather than a research session. It is the core concept of this RAG that everyone keeps talking about.
Why Hybrid is Even Better
Vector search is great at understanding intent. Keyword search is great at being literal. So, let's use both.
- Vector recall: “Startups using LLMs for regulatory intelligence” should surface teams that say “compliance co‑pilots for banks,” even if they don’t type “regulatory intelligence.”
- Keyword precision: If you require the literal word “e‑commerce” or a specific city, BM25 keeps the list honest.
Then a reranker takes the blended list and reorders it by how well each description answers the query in context. Think of it as the last‑mile quality check, precision over raw similarity, so the best‑fitting cards rise to the top.
A quick scenario:
Ask: “Affordable carbon accounting tooling for SMB e‑commerce brands.”
Vector pass: Finds “sustainability ledgers for Shopify merchants,” “Scope 3 tracking for DTC brands,” and similar neighbors in concept space.
BM25 pass: Nudges up cards that literally mention “e‑commerce" or “carbon.”
Reranker: Boosts longer descriptions that follow the same story arc, SMB focus, price sensitivity, and store integrations, so the top of the list feels tight.
Simple? Yes.
Founders and Investors Both Win
Founders get found for what they’re building, not only the buzzwords on their landing page. The system scrapes, normalizes, and enriches websites so even a stealthy team ends up with clean product descriptions.
Investors get nuance without babysitting the query builder. Combine conceptual searches with hard filters (funding ranges, geography, team size) and let AI push the noise out of the frame.
The Tech Stack (And Why I Like It)
- Voyage AI (embeddings & reranker). In my tests, Voyage’s latest embedding models deliver the best recall and precision by miles. Call it state of the art for practical retrieval, reliable, strong on intent, and easy to work with.
- Zilliz (managed vector database). Built by the folks behind the open-source Milvus, Zilliz gives you a managed vector store with ridiculously fast nearest‑neighbor search and sensible pricing.
- Supabase (the web’s standard toolkit). Auth, Postgres, storage, row‑level security, without drama. It’s the backbone for sessions, analytics, and a few admin tables. If you’ve shipped modern web apps lately, this will feel familiar.
- Next.js with Vercel is the de facto standard for a lot of web tools.
Where This Heads Next
I want to index more than startups. Think GitHub projects, open‑source libraries, developer tools, research repos, and public docs that founders and engineers actually use. If a great utility lives on page seven of search, it should still be easy to find when you describe what it does. The goal is one intent‑friendly place to discover teams and the tools they ship.