Building Intersection Intelligence

Building

Intersection
Intelligence

A personal mapping experiment
built with Codex and Claude

Mo Khan
14 November 2026

A question about a place

What is here,
and how do we know?

A personal research question that led me into an unfamiliar geospatial stack.

Start with South Africa

Research app screenshot with sensitive UI redacted

Country scale
Geographic context
Evidence with limits

Research app, 14 Nov 2026. UI redacted.
© OpenStreetMap contributors / Protomaps.

The journey in the repository

30 AUG

Standalone fixture MVP

31 AUG–1 SEP

Real geography and vector atlas

2–17 SEP

Contributions and stable identities

19–25 SEP

Measured performance work

LATE SEP

Contextual AI and shared platform

OCT–NOV

Country expansion and hardening

Zoom into Gauteng

Province scale
Bounded evidence
Clusters and points

Research app, 14 Nov 2026. UI redacted.
© OpenStreetMap contributors / Protomaps.

Find the city-scale question

Johannesburg
Local context
Details on demand

Research app, 14 Nov 2026. UI redacted.
© OpenStreetMap contributors / Protomaps.

Follow a corridor

City to corridor
Published road topology
A connected place

Research app, 14 Nov 2026. UI redacted.
© OpenStreetMap contributors / Protomaps.

One place, several kinds of evidence

OpenStreetMap

Geography
and recorded tags

Authority sources

Reported assets
and source dates

Human evidence

Observations
and review history

Area summaries

Counts at their
stated scope

Resolved view with provenance

Map, dossier, count and export use consistent scope.

Unknown, absent, unavailable and zero keep their own meanings.

Every source needed a way in

1. Retain source

Original fields, IDs and dates

2. Custom importer

Source-specific parsing

3. Candidate matching

Fuzzy names, road pairs, geography

4. Assisted review

App + browser location checks

5. Explicit acceptance

Candidate and confirmed differ

6. Publish and reconcile

Keep unresolved rows and history

Google Maps supported review, not unrestricted copying. Every source retains its terms.

Refreshes had to preserve human knowledge

Source revision A

Mapped nodes and roads

Stable place identity

Aliases and explicit reconciliation

Source revision B

Changed geometry or grouping

Preserved human knowledge

Contributions, accepted decisions and complete history

Illustrative model. A refresh must not detach accepted evidence.

Codex and Claude worked through evidence

Owner

Intent, scope and acceptance

Codex

Implement and verify

Claude

Challenge assumptions and inspect evidence

Curated ledger + tests + live results

This describes recorded collaboration sessions, not independent review of every commit.

Independent review changed the work

Build location

Heavy geography processing moved off the serving host.

Application scope

A map area could not inherit financial assumptions.

Completion evidence

Source review, test results and live acceptance stayed distinct.

Software architecture

Atlas and dossiers

JavaScript / MapLibre / PMTiles

Shared chat client

React conversation interface

Product API

Node.js / Fastify + domain code

Shared Signal AI

Product-owned read tools

Prepared reads

Files + read-only SQLite

Human evidence

Events + conditional heads

Platform identity

Canonical entitlements

The technology stack

Browser

JavaScript / MapLibre GL JS
PMTiles / Protomaps / React chat

Application

Node.js 24 / Fastify
Shared geography and evidence code

Preparation

Python geospatial processing
Custom importers / matching

Data

Versioned files / SQLite
Blob events + conditional heads

Hosting

Azure App Service Linux B2
Entra ID / Blob Storage

Shared services

Signal AI / entitlements
GitHub CI / release verification

A concrete stack for a personal research application. Resource identifiers omitted.

Building and serving are different workloads

BUILD

SERVE

Pinned sources
Python geospatial processing
Identity and spatial validation
Versioned artifacts

Interactive map
Bounded API reads
Local analytical workers
Attributable contributions

Heavy extraction runs outside the serving worker and ordinary application CI.

Deployment architecture

Entra ID

Enterprise sign-in

User browser

Atlas + shared chat

Basemap Storage

Signed tile reads

Release workflow

Code / validated data

App Service

Authenticated product API

Shared services

Entitlements + AI

Application storage

Immutable files + local SQLite

Evidence Storage

Separate history / restricted network

Service endpoint with restricted public Storage access. Identifiers omitted.

A small Linux host shaped the design

Linux B2

2 cores · 3.5 GB RAM

One instance / Node.js 24 LTS / Always On

Prepare geography elsewhere. Serve bounded reads.
Recover local SQLite copies from verified persistent releases.

B2 lists 10 GB plan storage; temporary worker disk is separate. No sustained capacity claim.

A map release must preserve contribution history

IMMUTABLE RELEASE

HUMAN EVIDENCE

Prepared geography
Versioned sources
Read models and indexes

Create-only events
Conditional current heads
Committed ancestry

Stale edits conflict. Map refreshes preserve accepted history.

Rendering improved through repeated cycles

Draw the useful view

Bound visible objects, cluster points and reuse coordinates.

Stabilise the camera

Reserve layout space and query the final map bounds.

Reuse fresh evidence

Keep complete context keys, expiry and edit invalidation.

Measure the visible result, keep the caveats

30 cached revisits

Zero requests and aborts; same visible evidence.

90.5 ms median

116.5 ms p95 application completion; GPU paint excluded.

771 to 529 ms

Street API median improved; overview median regressed.

A measured improvement in repeated queries

855 ms

276 ms

Repeated count request
Median of three warm repeats

22 September production experiment. First request essentially unchanged. Not a whole-app or capacity claim.

Performance meant doing less repeated work

Select scope

Use the requested area before expensive enrichment.

Reuse evidence

Validate freshness and retain unchanged prepared work.

Preserve answers

Check counts, dossiers and exports against the same population.

AI in the product uses authorised evidence

Selected map context

Area, record and original turn scope

Authorised read tools

Bounded product evidence

Deterministic facts

Counts, calculations and provenance

Conversational answer

Explicit gaps, partial results and uncertainty

Development AI helps build the system. Product AI reads evidence through its own contracts.

The first usable release required an owner decision

Implementation

Code and regression evidence

Deployment

Exact running application and data

Acceptance

A usable experience with stated limits

The first native AI increment launched at the end of September under a bounded owner decision.

The accepted research slice

AVAILABLE

STILL OPEN

Atlas and site dossiers
Source-aware comparisons
Reviewed contributions
Contextual AI
Nine-country catalogue

Uneven source depth
Sustained capacity acceptance
Richer topology and calibration
Gated data-maintenance procedure

Discovery-and-capture alpha. No live traffic-control or validated traffic-benefit claim.

Original vision, accepted MVP

Explore a place

Accepted: country, city, junction and corridor navigation.

Trust the evidence

Accepted: source-aware dossiers, candidate links and reviewed edits.

Model the wider network

Open: complete coverage, calibration and sustained capacity.

What I would repeat

A narrow real-data slice

It reveals the work hidden by a convincing prototype.

Identity and provenance early

People need to trust what survives the next refresh.

Specific review questions

Findings need evidence and an implementation owner.

User-visible acceptance

A green build is one milestone on the way to a useful product.

Learning a new stack with AI

A useful research MVP, a clearer understanding,
and a longer list of questions worth exploring.

Mo Khan

Questions and discussion