One engine. Many products.
A single deterministic, vector-less engine for structured data. It finds records that are similar to a given one, records that are statistically out of place, and the patterns that repeat across your data — directly on your raw fields. No vectors. No models. No pipeline.
What it is
Pop is a deterministic, vector-less engine for structured data. It works on the numeric and categorical fields you already have — price, rating, latency, duration, revenue — with no embedding models and no black box.
The same core does three jobs: find records that are similar to a given one, find records that are statistically out of place, and find the patterns that repeat across your data. Every result comes with a score you can read and a reason you can explain, which matters when the answer has to stand up in a pricing review, a fraud queue, or a product decision.
That one core powers different products for different jobs — a search API for your catalog, an analytics engine for your web events, and more on the way.
Products
Each product is the same engine pointed at a different job. Each runs as its own single binary.
Structured similarity & anomaly engine. Load a catalog as plain records and get similar records and statistical outliers back, on your raw numeric fields.
Open Pop-Search →Web analytics — your website, understood. Instant audience segments, in-session conversion funnels, fraud & bot detection, and journey patterns over your web events.
Open Web-Pop →DOCX to PDF / EPUB conversion. Engine layout algorithm, font embedding, tables, images, footnotes. Upload, convert, download — nothing stored.
Open Doc-Pop →The same engine, pointed at new jobs. Same core, new data — more products are on the way.
How it works
Every product shares the same three steps. You load your data; the engine indexes your fields; you query over HTTP — and it stays fast on a small 2-core / 4 GB box.
Load your data as plain records — JSON, CSV, or a stream of events.
The engine indexes your numeric and categorical fields for fast similarity, range, anomaly, and pattern lookups.
Ask over HTTP and get interpretable results back — scores you can read and explain.
We can't share the internals just yet. The short version: it's deterministic, interpretable, and vector-less — and it's built to stay fast on modest hardware. This is a beta, open for public testing of the algorithm. What you see here is what we can show today; the details are on the way.
Why it's different
The tools people reach for today were built for a different job.
| Vector DB (Qdrant / Milvus) | Search engine (Elastic / Meili) | Postgres / SQL | Pop | |
|---|---|---|---|---|
| Needs embeddings? | Yes | pgvector: yes | pgvector: yes | No |
| Needs a model / GPU | Yes | Optional | Optional | No |
| Numeric-field similarity | Via vectors | No | Hand-rolled SQL | Native |
| Anomaly / outlier detection | No | No | Hand-rolled | Native |
| Recurring pattern mining | No | No | Hand-rolled | Native |
| Interpretable results | Black box | Text relevance | You define | Yes |
| Runs on 2-core / 4 GB | Hard | Yes | Yes | Yes |
Hardware
Each product is one static binary with local file persistence. It runs comfortably on a small VPS — the kind of box that hosts a dozen little services — with no JVM, no cluster, no external database, and no GPU.
That's the footprint per product — and there's still room for the rest of the stack.
Feedback