Hackney Economic Atlas (kaomapi-2026)

A high-dimensional, high-granularity open-data atlas of the London Borough of Hackney (E09000012), built to power agentic systems for economic decision-making — helping small businesses pick venues, renters analyse areas, and councils target intervention, by modelling how money flows across businesses, workers, and markets. Designed for local processing on an NVIDIA DGX Spark.

Status: built (2026-06-06). Private dataset repo. 19 data domains + spine, ~151 artifacts, each with provenance + QC + a per-dataset DATA_DICTIONARY.md. The 33 GB of raw national source bulks are archived under raw/. A few layers await one-time user action (see "Pending" below).

Layout

spine/        geography & identity backbone (boundary, wards, OA/LSOA/MSOA21, ONSPD, OS UPRN/roads, OSM)
curated/      domain = businesses | property | mobility | people | environment | publicmoney | visual
              | heritage | water | health | education | economy | transport | housing | buildings
              | costs | consumer | infrastructure | corpus | datastore
              keyed to (uprn?/lat/lon/geo_code, period, measure, value, source_id)
manifest.parquet   one row per artifact: path, domain, rows, bytes, sha256, source_url, licence, public
raw/          the source national bulks (ONSPD, OS UPRN/Roads, OSM, Companies House, EPC, Land
              Registry Price Paid, crime/STATS19, Colouring, Sentinel-2, LIDAR …) for re-derivation

Domains of note: costs = regulatory/tax/fee layer (business rates, CIL, council tax, ULEZ, NLW/VAT/CT, utilities) for P&L and cost-of-living modelling; consumer = CDRC engineered indices (IUC, AHAH); infrastructure = heat networks, mobile coverage, public toilets, imagery catalogue; corpus = RAG text (Local Plan + London Plan policy, licensing & planning committee reports) plus webcast_catalogue.parquet — the NeMo-ASR worklist of council-meeting recordings to transcribe; datastore = Hackney's slices of the wider London Datastore (wellbeing/zen index, rents, measured health, demographics, creative geography, and the cross-borough analysis leads) — each value carries its London rank/percentile for context (e.g. emergency admissions 33/33 lowest while deprivation 1/33 highest).

Each domain folder also ships provenance.json, QC_REPORT.md, and DATA_DICTIONARY.md (a per-dataset column dictionary). Collector scripts live in the git repo (not here). Live-feed frames are collected by cron (see git docs/CRON_TODO.md).

Pending (one-time user action — not yet ingested)

• CCOD/OCOD property ownership — API key works, both free, but the dataset licence must be signed once at use-land-property-data.service.gov.uk. Then the collector path runs.
• EGMS InSAR ground motion/subsidence — needs a free EU Login (catalogued in infrastructure/imagery_catalogue.parquet).
• Skipped (documented): HMO register (pay-walled portal), Shopfront Design Guide SPD (scanned, no text layer — needs OCR).

Usage — loading from Hugging Face

Every artifact is parquet, readable directly over hf:// — no download step.

DuckDB (recommended for SQL/joins):

import duckdb
con = duckdb.connect()
con.sql("INSTALL httpfs; LOAD httpfs;")
REPO = "hf://datasets/kaomapi/kaomapi-2026"
con.sql(f"SELECT * FROM '{REPO}/manifest.parquet'")                       # what's in the atlas
con.sql(f"SELECT * FROM '{REPO}/curated/businesses/voa_rating_list.parquet' LIMIT 5")
# join any layer to the geography spine (UPRN address grain):
con.sql(f"""
  SELECT u.uprn, v.rateable_value, f.business_name
  FROM '{REPO}/spine/uprn_hackney.parquet' u
  JOIN '{REPO}/curated/businesses/voa_rating_list.parquet' v USING (uprn)
  LEFT JOIN '{REPO}/curated/businesses/fsa_hygiene.parquet' f USING (uprn)
  LIMIT 20
""")

pandas / polars:

import pandas as pd
df = pd.read_parquet("hf://datasets/kaomapi/kaomapi-2026/curated/people/claimant_count.parquet")

cuDF (RAPIDS, on the DGX Spark) — same paths, GPU dataframes:

import cudf
gdf = cudf.read_parquet("hf://datasets/kaomapi/kaomapi-2026/curated/mobility/hsp_delays.parquet")

huggingface_hub (bulk pull):

from huggingface_hub import snapshot_download
snapshot_download("kaomapi/kaomapi-2026", repo_type="dataset", local_dir="atlas")

Start from manifest.parquet to discover every layer, its grain, source and licence. The geography join key is ONS 2021 codes (oa21cd/lsoa21cd/msoa21cd/osward) or uprn at address grain (in spine/uprn_hackney.parquet).

Provenance & licensing

Each domain ships a provenance.json and QC_REPORT.md. Sources are OGL v3.0 / CC-BY unless noted. Some layers are aggregated rather than raw where the source licence restricts redistribution (e.g. EPC certificates, Land Registry CCOD/OCOD ownership). Nothing person-level is published beyond what the original source already publishes openly.

Geography

The join backbone is ONS 2021 geography (OA → LSOA → MSOA → ward → borough) plus OS Open UPRN at address grain. Built with the tooling in the companion analysis repo.

---

Hackney spans ~280k residents, ~19 km², 21 wards, ~149 LSOAs, ~795 OAs, ~184k UPRNs.