Marti Soura Vamseekar

Work / Aequitas

Aequitas

Answers one question the same way in every country: who here is underserved by public transport, and by how much?

LiveOct 2021 – Present
aequitas.souravamseekar.comliveSVamseekar/aequitas

The problem

Your country already publishes the pieces: timetables, small-area geography, and an official deprivation index. What it does not have is a briefing that turns them into the same answer every time. So each region commissions its own study, each study picks its own method, and none of the findings can be set beside another — including the ones from the region next door.

Finding

People within 400 m of a stop · by deprivation decile0%25%50%75%100%12345678910most deprivedleast deprivedEnglandIMD 2025400 m · 79.3%score 80.0IrelandPobal HP 2022400 m · 55.1%score 55.5NetherlandsCBS SES-WOA400 m · 91.9%score 69.6FranceF-EDI 2021400 m · —score 47.7Each country is scored against its own index — these numbers never share an axis.
Four countries, one question. In every one, the least deprived are best served — and each is scored against its own national index.

A grouped bar chart of transport coverage by deprivation decile for the four live countries. In every country coverage rises from the most deprived decile to the least deprived. National figures from the current warehouses: England 79.3 percent coverage within 400 metres of a stop, scoring 80.0 against IMD 2025; Ireland 55.1 percent, scoring 55.5 against Pobal HP 2022; the Netherlands 91.9 percent, scoring 69.6 against CBS SES-WOA; France scoring 47.7 against F-EDI 2021. The scores are computed inside each country and are not comparable across them, because each deprivation index is constructed differently.

How it works

  1. 01

    It starts from what you already publish

    Your national timetable feed, your census geography, your deprivation index. No proprietary data, no invented travel times, nothing you cannot audit — England runs on IMD, Ireland on Pobal HP, the Netherlands on CBS SES-WOA, France on F-EDI.

  2. 02

    Your score is yours

    The same formula runs inside each country against that country's own index. Your regions are ranked against each other, never against another country's — because your index and theirs do not measure the same thing.

  3. 03

    You get a briefing, not a dataset

    A map, a quotable national score, and a fixed set of exhibits: who lives near a stop, who is stranded after seven, how service quality tracks deprivation across your own deciles.

  4. 04

    It travels to your country next

    Adding a country means supplying its feed, its geography and its index — the analytics, the validation gates and the briefing come with the method. Four are live; the fifth is a data question, not a rebuild.

Engineering notes

  • Four national warehouses live, each with its own ingestion, processing and banding modules: England (IMD 2025, LSOA 2021, 79.3% within 400 m, score 80.0), Ireland (Pobal HP 2022, 18,919 small areas, 55.1%, score 55.5), the Netherlands (CBS SES-WOA 2023, 91.9%, score 69.6 on bus) and France (F-EDI 2021, IGN IRIS, 48,522 areas, score 47.7).
  • Analytics are pre-computed when the warehouse is built; the API is a lookup layer, so the interface cannot invent a figure at request time.
  • Composite score weights 400 m coverage, evening service, weekday quality and the deprivation–service correlation, renormalising when a term is unavailable.
  • Thirteen product surfaces including equity, access, service, network HHI, correlations, scenarios, time series and in-country compare.
  • Narratives are templated from warehouse numbers; weak evidence is dropped rather than imputed.

Pipeline

Run once per country — sources and index differEnglandIrelandNetherlandsFranceIngestGTFS · censusfeeds · boundariesProcessjoins · dedupgeometry · frequencyAnalyticsequity · access2SFCA · clusteringValidategates · truth103 checks · 0 failWarehousepre-computepacks · provenanceDuckDB warehouseevery section pre-computedRead-only APIlookup, not computeDashboard13 surfaces · mapsGrounded chatFAISS index over sectionsScores are computed inside a country.Indices are constructed differently, so theynever share an axis and never form a league table.Weak evidence is omitted, never imputed.
One pipeline, run once per country. Each nation brings its own sources and deprivation index; the warehouse it produces is queried, never recomputed.

Aequitas runs the same five-stage pipeline separately for each of four countries. Ingest reads official GTFS timetables, census geography and the national deprivation index. Process performs spatial joins, deduplication, route geometry and service frequency. Analytics computes equity, accessibility including 2SFCA, economic bands, and machine learning modules for clustering, anomaly detection and prediction. Validate runs quality gates and ground-truth checks. Warehouse pre-computes every analytical section with provenance. England, Ireland, the Netherlands and France each have their own ingestion, processing and banding modules. A read-only API serves the pre-computed warehouse to the dashboard and a retrieval-augmented chat index.

Trade-offs

In-country scores only

instead of one European league table

IMD, Pobal HP, SES-WOA and F-EDI are constructed differently. A table ranking your country against another would look authoritative and mean nothing — the most requested feature and the wrong one to build.

Pre-computed warehouse

instead of computing analytics per request

It makes the API a lookup layer, which is both fast and honest: nothing can be estimated on the fly to fill a gap.

Omitting weak exhibits

instead of imputing missing values

An empty panel tells a transport authority the truth. An imputed one invites a decision the data cannot support.

Evidence

4

national warehouses live

103

quality checks, 0 failures

13

product surfaces per country

1

method, applied per country