Wombie Intelligence

AI that reads the real world.

Every day, thousands of events surface across hundreds of fragmented, multilingual, half-duplicated sources. Wombie runs them through a machine-learning pipeline that understands what each one is, merges the duplicates, and turns the noise into a single, structured map of where to be tonight.

stand-up · tonight live techno — warehouse jazz @ blue note rooftop social · fri
The thesis

There is no API for reality. So we built one.

A concert, a trivia night, a film screening, a street festival — every source publishes them differently, in different languages, with different fields, half of them duplicated three times over. The real world does not arrive pre-structured.

Rules and scrapers alone fall apart at this scale. Reading messy human text and deciding what it means is exactly what modern AI is for — so machine-learning models sit at the center of everything Wombie does, from the first byte ingested to the recommendation that reaches you.

01 — The Engine

Noise in. Knowledge out.

Every event flows through three stages. The middle one — understanding — is pure machine learning, and it is where the magic happens.

Ingest
Raw, messy, multilingual.

Hundreds of sources, scraped continuously. Inconsistent formats, missing fields, listings in a dozen languages, duplicates everywhere.

Understand · AI
The models read it.

Transformer-based models classify each event, resolve duplicates into one, enrich it with structure, and score its quality — no hand-written rules required.

classify de-duplicate enrich score
Structure
One clean map.

A single, de-duplicated, categorized catalog of real-world events — searchable, rankable, and ready to put people in the same room.

02 — Capabilities

Six places AI does the work.

Machine learning is not a feature bolted onto Wombie. It is the substrate the whole product runs on.

Zero-shot classification

Our models understand an event from a single sentence — concert, theatre, trivia, screening — with no labelled training data and across languages. New categories cost a string, not a dataset.

AI entity resolution

The same event shows up on ten sites under ten names. Language-aware fuzzy matching and semantic similarity collapse them into one canonical record — so you see the night, not the noise.

Intelligent enrichment

Models infer category, link screenings to a film catalog, normalize venues and places, and score completeness — turning a bare listing into a rich, trustworthy entry.

AI for community health

We are extending the same natural-language models inward: reading the tone of group chats to catch last-minute decline cascades and nudge gently, so plans survive contact with real life.

Semantic search & recs

Vector embeddings let people search by intent — "something low-key on a Tuesday" — and power recommendations built on co-attendance, not engagement-bait ranking.

AI-native engineering

A small team ships at the pace of a large one because the system that builds the system is itself AI. New sources, new models, new surfaces — generated, reviewed, shipped.

03 — Under the Hood

Real models. Real inference. In production.

Wombie serves transformer models natively from its Elixir backend — compiled, fast, and built to scale onto cloud GPUs as we grow.

Model card · in production
live
bart-large-mnli
Transformer · zero-shot NLI classifier
Task
Zero-shot classification
Labels
Open-set, no retraining
Inference
XLA-compiled, GPU-ready
Reach
Multilingual scaling
# a sentence in, a category out
classify("Pub quiz, every Thursday 8pm")
#=> label: "trivia" · score: 0.93
Bumblebee + Hugging Face

Pre-trained transformers loaded straight from the Hugging Face Hub and served in-app.

Nx + EXLA

Tensor compute compiled through XLA for fast CPU inference today, GPU inference as we scale.

Cloud GPU inference

A serving path built to lift onto managed GPU infrastructure as event volume and model count grow.

Vector embeddings

Semantic representations of every event powering search-by-intent and similarity matching.

At scale
1000s
events read & structured by AI
100s
fragmented sources unified
20+
countries, many languages
0
labelled datasets required
04 — What's Next

Deeper models, more of the world.

Every direction below leans harder on AI — and on the cloud GPU capacity to run it.

01

Embeddings-powered semantic search

Find events by vibe and intent, not keywords — at scale, across languages.

02

Co-attendance recommendations

Models that suggest who and what to do next from the real graph of who showed up together.

03

LLM-assisted enrichment

Large language models filling gaps, summarizing, and translating listings into one clean shape.

04

Managed GPU model serving

Bigger, multilingual models served from cloud GPUs as event volume and model count climb.

Building the intelligence layer for real-world connection.

AI is how we read a chaotic world and hand you somewhere to be. We are just getting started — and we are looking for partners, infrastructure, and people who want to build it.