ES
← Back to Portfolio
Data Engineering June 2026

Atalaya — A Watchtower over Chile's Open Data

Atalaya harvests Chile's Data Observatory open catalog, profiles every downloadable table, and mines five kinds of cross-dataset relation — same-source, semantic similarity, spatial overlap, joinability and statistical correlation — into an explorable knowledge graph, with client-side semantic search running in the browser. Honest about evidence strength, never causal.

Pipeline
harvest → profile → relate → evaluate → export (1,017 datasets, multi-GB partial mirror)
Relations mined
same-source · semantic (MiniLM/ONNX) · spatial · joinable (MinHash, 117) · correlated (Spearman+FDR, 24)
Model ladder
classical (PCA/KMeans + TF-IDF foil) + SOTA (MiniLM embeddings) + a novel calibrated multi-evidence affinity score
Honesty
Not causal; SOTA beats classical only ~+1.4 pts; small-n correlations flagged fragile
Stack
Python pipeline → React SPA · transformers.js / onnxruntime-web · KaTeX · static, no backend
Atalaya — A Watchtower over Chile's Open Data — Architecture
#data-engineering #knowledge-graph #information-retrieval #embeddings #onnx #open-data

Business Context

Public data is only as useful as your ability to connect it, and the connective tissue — which datasets join, overlap or correlate — is exactly what catalogs never provide. Atalaya turns a flat catalog into a navigable graph of relationships, which is the difference between "here are a thousand tables" and "here are the two that, joined, answer your question." Doing it with statistical discipline (nulls, FDR, partial-correlation guards) is what keeps the connections honest.

Strategic Value

Atalaya demonstrates an end-to-end data-engineering + IR stack — catalog harvesting, profiling, entity-resolution/joinability, statistically-vetted relations and a calibrated fusion score — delivered as a static, backendless, client-inference web app. Its discipline is the point: it reports evidence strength honestly (of ~14,000 edges, the vast majority are cheap priors; the hard evidence is a few hundred joins and a handful of FDR-controlled correlations), it never implies causation, and the modern embedding model beats the classical baseline only modestly. That honest, reproducible framing is what makes it a credible relation explorer rather than a graph that overstates itself.

The Challenge

An open-data catalog is a list of tables, not a map of how they relate. Chile's Data Observatory publishes over a thousand public datasets, but which two could be joined, which cover the same territory, which correlate, which come from the same source — the questions that make open data actually useful — are invisible. Answering them at catalog scale needs real data engineering (harvest, profile, entity-resolve) and statistical care, not a keyword search.

Our Approach

Atalaya is a data-engineering + information-retrieval pipeline — harvest → profile → relate → evaluate → export — run over 1,017 real datasets from the catalog (with a measured multi-GB partial mirror). It profiles every table, then mines five relation types into a knowledge graph: same-source, semantic similarity (a MiniLM sentence model exported to ONNX), spatial overlap, joinability (MinHash containment) and statistical correlation (Spearman with a permutation null, Benjamini-Hochberg FDR control and a partial-correlation guard). A novel calibrated multi-evidence "affinity" score fuses the signals against null-distribution percentiles and reliability weights, and can be re-weighted live. It ships as a static React SPA with the graph baked in and semantic search running client-side (transformers.js / onnxruntime-web) — no backend.

Key Performance Indicators

KPIBaselineResultImpact
From catalog to mapA flat list of tablesA knowledge graph of 5 relation types over 1,017 datasetsFind the datasets that actually connect
Honest evidence"14,000 relationships"Strength labelled: cheap priors vs 117 joins + 24 FDR-controlled correlationsYou know which edges to trust
Where it runsA served backendStatic SPA; semantic search runs client-side (ONNX)No server, fully inspectable

Architecture

atalaya graph

atalaya graph

A map of how open data connects

Atalaya is a watchtower over Chile’s open data. It harvests the Data Observatory catalog, profiles every downloadable table, and mines five kinds of cross-dataset relation into an explorable knowledge graph — turning a flat list of a thousand-plus datasets into a map of which ones join, overlap, correlate or share a source. Live at atalaya.fasl-work.com.

Five relations, mined with care

Over 1,017 real datasets: same-source, semantic similarity (a MiniLM model exported to ONNX), spatial overlap, joinability (MinHash containment) and statistical correlation (Spearman with a permutation null, Benjamini-Hochberg FDR control and a partial-correlation guard). A novel calibrated affinity score fuses the signals against null-distribution percentiles and can be re-weighted live. The whole thing ships as a static SPA with the graph baked in and semantic search running client-side — no backend.

Honest about the graph

The number that matters is not “14,000 relationships” — most of those are cheap priors. The hard evidence is a few hundred joinable pairs and a handful of FDR-controlled correlations, and Atalaya labels that strength rather than hiding it. It never implies causation (some small-n correlations are flagged fragile), and the modern embedding model beats the classical TF-IDF baseline only modestly (+1.4 points). It is a relation explorer, reported at the confidence the data supports.

Live demo · Source on GitHub

Technology Stack

Python NumPy MiniLM ONNX React TypeScript

Visual assets for this project are not publicly available.