🕵️ Dive Bar Detective

Want your city supported?

I'd love to expand this to more cities! However, current data acquisition and API costs are higher than I'd like to do this work at scale.

Interested? Let me know!

✓ Request submitted! Check your email for next steps.

This project has real costs

Running Dive Bar Detective involves ongoing expenses:

Buy me a coffee (or beer!) 🍺

☕ $5 — Coffee 🍺 $10 — Beer 🍽️ $20 — Lunch

What this is

A data-driven "Sommelier for Dive Bars" that helps you find off-the-beaten-path spots in your town (or for now, Denver only). We use ML, AI and NLP to analyze reviews and extract what matters (at least what matters to Jeff): food quality, vibe, authenticity, and local recommendations.

The dataset

251 locations and 4,260+ reviews from Denver bars and restaurants. Being this is a proof of concept, I didn't pull in all locations from Google's Places API, but rather selected a subset of locations that I thought would be interesting and would be a good fit for the project. Overall, I wanted to select a diverse set of locations that would be a good fit for the project. No chains, focused on neighborhood joints with real character. Each review is analyzed to extract 9 quality and vibe signals..

How scoring works

Scores use percentile-based ranking within our dataset. A score of 9.5 means "top of our curated list" — not a perfect 10, but the best relative to other authentic Denver spots. This makes scores comparable across all lenses and helps you quickly identify standouts. And again, the dataset was already limited to only 251 locations, so the scores are not perfect, but they are a good starting point.

The lenses (0–10)

Two ways to explore

How the Vibe Map works

The Vibe Map visualizes all 251 locations based on their Quality and Character scores. Dashed lines show the median — places in the top-right quadrant are above average on both axes. Bubble size reflects the current lens score. Scroll to zoom, shift+drag to pan.

Tech behind the scenes

NLP Pipeline: LLM-based sentiment analysis and signal extraction from review text.
UMAP (umap-learn): Dimensionality reduction for clustering similar places.
scikit-learn: K-Means clustering for vibe tags, Isolation Forest for "Unique" badge detection.
BERTopic: Auto-tag extraction (e.g., "late-night", "craft-beer") from review topics.
Transformers: RoBERTa-based sentiment scoring as a secondary signal.