An interactive, single-file analytics dashboard over 1,001 container terminals, rail intermodal yards and industrial sites worldwide — modelling how gate/crane/rail OCR vendors are distributed across the global port universe, where the whitespace is, and when incumbent contracts come up for renewal.
No build step, no server, no framework. One HTML file, ~1.5 MB, opens straight from disk.
This is a de-identified, partly synthetic derivative of a private market map. Two different things are going on in the same file:
| Real | Site coordinates (public container terminals, rail yards, industrial sites), the facility taxonomy, the region split, and the Confirmed / Inferred / Conflict / TBD attribution grading. |
| Anonymised | Every vendor, operator and site name. Vendors are Company 1…15, operators continue the same identifier pool, sites are Site <id>. All source references, evidence text and URLs were dropped. |
| Modelled | Every operational, contractual and commercial attribute — throughput, equipment counts, OCR accuracy, contract dates, licence value, risk scores. These are generated, not measured. |
Nothing here is a fact about any real port, operator or vendor. What it does demonstrate is a data model, an analysis layer, and a rendering approach.
The app says so itself: a disclosure banner on load, and a Methodology tab with a
per-field provenance table tagging every field observed / anonymised / modelled.
The private version carries exactly one attributed field per site: which vendor's system is installed. That's enough to draw a footprint map and not much else.
This build asks a different question — what would a complete per-terminal intelligence record look like, and what analysis would it unlock? — so it carries the full 66-field schema and fills it from a model.
The generator is deterministic: every value is a hash of the site id and the field name. Builds are reproducible and diffable, and the fields stay consistent with each other rather than being independently random:
- crane and berth counts follow throughput
- truck turn time follows gate automation and lane count
- OCR accuracy follows vendor tier and upgrade recency
- licence value follows throughput, channel breadth and sales channel
- displacement risk is a real formula over contract proximity, accuracy gap, integration depth, support tier and upgrade age
The distributions are plausible. The individual values are fiction.
Seven tabs, all driven by one sticky filter bar (region · facility type · attribution status · automation level · vendor · size class · full-text search), plus a metric switch that repoints every chart between site count, throughput and licence value.
Eight KPIs (covered volume, installed licence base, open opportunity, contracts expiring ≤24 months, HHI concentration), vendor share, vendor mix by region with a 100%-stacked toggle, an adoption timeline stacked by vendor with annual/cumulative toggle, attribution-quality and freshness donuts, and a sortable top-N table with inline throughput sparklines.
Sortable 12-column leaderboard, then a profile for the selected vendor: modelled identity card, regional footprint, OCR channels held, install cohort, contract runway, largest sites — plus a head-to-head comparison against any rival with directionally-correct relative bars (lower truck turn and lower risk score as wins).
A vendor × OCR-channel heat matrix (channels held or throughput covered), an interactive scatter with five selectable Y-axes and four colourings (bubble size = licence value, click any point to open the record), automation posture by region, mean accuracy by incumbency band, feature adoption, and an OCR-vendor × TOS crosstab.
Weighted pipeline value, renewal runway by year stacked by incumbent, a region ×
years-to-renewal pressure heatmap, unattributed value by region, a qualification
funnel, and a ranked target list scoring whitespace wins and incumbent displacements
on one comparable axis (deal value × probability × timing).
All 66 fields as 57 selectable columns, grouped column chooser, sort on any column, pagination, and CSV export of either the visible view or the complete field set.
Leaflet map of all 1,001 sites. Marker shape = facility type, size = throughput, fill = attribution status, outline = engagement type (direct / reseller / source conflict / web-inferred). Optionally inherits the dashboard filters.
Field-by-field provenance, coverage stats, vendor taxonomy, and definitions.
Click any row, bar or point anywhere in the app to open a full 66-field record drawer.
66 fields per terminal.
| Group | Fields |
|---|---|
| Identity | id name region segment type size_class operator multi_op lat lon |
| Attribution | vendor status mm_claim df_claim confidence_score data_completeness internet_inferred confidence last_verified freshness |
| Capacity | volume volume_unit vol_history[5] cagr_3y berths quay_m depth_m yard_ha annual_trucks_k |
| Equipment | sts_cranes yard_cranes reach_stackers gate_lanes reefer_plugs on_dock_rail rail_tracks |
| OCR stack | ocr_stack{Gate,Crane,Rail,Yard/UTR → vendor + year} ocr_channels channel_count multi_vendor |
| Systems | tos integration hosting automation gate_automation_pct anpr damage_ai weighbridge_link |
| Performance | ocr_accuracy uptime_pct truck_turn_min gate_moves_hr |
| Commercial | acv_kusd tam_kusd sales_channel support_tier install_year last_upgrade_year contract_years contract_end refresh_due_yrs incumbency_yrs displacement_risk deal_stage |
Two modelling decisions worth calling out, because they're what make the analysis work:
OCR channels are drawn independently. A site can run different vendors on its gate, crane, rail and yard channels. That's what produces the 152 multi-vendor sites — and it's the wedge the Technology tab is built to expose.
Whitespace has no licence value. acv_kusd is null for unattributed and in-house
sites rather than defaulting to a number, so installed-base totals can't be silently
inflated by sites nobody has sold to. Whitespace carries tam_kusd instead.
data/terminals_anonymized.json is the same dataset as a standalone file if you want
it without the UI.
index2.html (private)
|
+- 1. ANONYMISE vendors + operators -> "Company N", sites -> "Site <id>",
| all free text, sources, evidence and URLs dropped.
| Coordinates kept - geography is public.
|
+- 2. EXTEND +234 candidate public terminal coordinates, de-duplicated
| against the base universe at 8 km great-circle distance.
| 155 survived. Vendor/status drawn from region-specific priors.
|
+- 3. ENRICH Each record expanded to the full 66-field profile from a
| deterministic hash seed.
|
+- 4. ASSEMBLE Emit one self-contained HTML file from src/tpl3/.
Step 4 rebuilds the page from templates in this repo rather than patching the private source, so there's no path for prose from the private file to survive into the public one by accident. Verified by a token-set diff of the two files' vocabularies: zero company, operator, port or city names, zero URLs, zero file paths.
python src/build_index3.pyWon't run from this repo — it needs the private
index2.htmland a git-ignoredprivate/identity_map.pyholding the only two mappings that name real entities. The script is included because the pipeline is the interesting part; the shippedindex.htmlis its output.
index.html the app - open it directly, no server needed
data/
terminals_anonymized.json the dataset on its own
src/
build_index3.py anonymise -> extend -> enrich -> assemble
tpl3/
ports_seed.py additional public terminal coordinates
enrich.py the deterministic attribute model
styles.css body.html page shell
core.js state, filtering, formatting, record drawer, CSV
charts.js SVG chart primitives (no charting library)
dashboard.js vendors.js technology.js pipeline.js database.js
map.js methodology.js boot.js
- No framework, no build tooling, no charting library. Every bar, column, donut,
scatter, heatmap, funnel and sparkline is hand-rolled SVG in
charts.js(~14 KB). - Only external dependency is Leaflet 1.9.4 plus CARTO basemap tiles, and only the Map tab needs them — the five analytical tabs work fully offline.
- Charts are strings; interactivity is delegated via
data-tip/data-actattributes and two document-level listeners, so re-rendering a whole tab needs no teardown. - Views render lazily and mark themselves dirty on filter change.
- ~2,000 lines of JS, ~600 of Python.
Not yet specified — add one before reuse.