GIFT City Twin Blueprint

The problem

About 25,000 to 30,000 people work in GIFT City. Their day is shaped by heat over 40 °C, a metro branch that ran only at peak hours, and almost no autos after 7:30 pm. The city's rich operational data never reaches them.

Our answer

One twin core with two faces: a phone companion for the people who walk, commute and work there, and a 3D console where operators and planners see forecasts and test what-ifs.

Proof so far

A working twin built in two days from open data: 1,718 buildings fused from five sources, live weather and air quality, a moving metro, and three what-if questions answered with numbers.

What we need

Four decisions (endgame, primary user, team and budget, GIFTCL access) and a short round of user interviews before the MVP is locked.

3.68 km²working boundary, the OSM place polygon

1,718building footprints fused, each with a confidence score

99cited sources, each tagged by confidence

13architecture decisions recorded as ADRs

153 → 7minutes average wait for a train home, if the metro ran every 15 min (modelled)

Why this twin, and why now

GIFT City already runs one of India's most instrumented utility networks, and an official portal twin is being built. Neither helps a commuter at 1 pm in May. We build the part that is missing: daily usefulness and prediction, on open standards so it can plug into the official platform later.

Documented pain	What people report	What only a twin can do about it	Fit
Heat on foot	Summers around 40 °C make walks impractical; staff ask for shaded paths.	Shade-aware walking and street-level heat stress need 3D buildings, sun position and weather together.	Strong
Getting home	Metro branch ran 07:36–10:18 and 16:06–19:13, every 49–57 min (Jan 2026). Autos scarce after 7:30 pm.	Door-to-desk planning across metro, GIFT Go and walking, with a warning before the last option leaves.	Strong
Finding the place	Visitors and new joiners cannot find the right tower, gate or lobby.	Building, entrance and company finder down to the lobby, with drop-off points.	Medium
After-work life	Described as a "ghost town after working hours".	Amenity and event discovery. Useful, but a map does this too.	Weak

Sources: Outlook Business 2026 (S04), Gujarat metro timetable reporting Jan 2026 (S12). No primary user research yet; see section 07.

Where we sit next to the official twin

	Official (expected)	This project
Buyer, vendor	GIFTCL, DEV IT; ₹2.79 crore, filed 1 Jul 2026	Independent; seeks partnership
Focus	Web portal, content, walkthrough, investors	Daily utility for people on the ground
Mode	Descriptive: what exists	Predictive: what will happen, what if
Openness	Vendor platform	NGSI-LD, CityJSON, 3D Tiles, MCP

What Indian city twins look like today

Varanasi, Ayodhya, Guwahati and, since September 2026, Ahmedabad (₹283 crore, 625 km², underground utilities) were all procured from mapping vendors. They are strong on base maps, CCTV and asset layers, and weak on prediction, what-if analysis and anything a citizen uses daily.

Our position: base mapping is a vendor game we should not play. The application and simulation layer on top, built on the same open standards, is open ground.

Complement, never duplicateIntegrate through standardsNever look official

Five principles we hold ourselves to

Useful on day one	Nothing in the public app may depend on data we do not have yet. Open data first; partner feeds plug in later.
Predictive, not descriptive	The MVP predicts at least street heat stress and journey reliability, and answers at least one what-if.
Honest by design	Every value shows how it was made. measuredderivedmodelledsynthetic
Built to integrate	Standards-based core with adapter interfaces, so GIFTCL's C4 or the official twin can connect without a rewrite.
Works in the sun	Companion interactive in 3 s on a mid-range Android over 4G; console at 30 fps on a mid-range laptop.

Who it serves

Five groups touch GIFT City every day. Each one gets something no map, portal or control-room screen gives them today. The first column is the people we would build for first.

Workforce and commuters

~25,000–30,000, incl. security, housekeeping and drivers

"It is 6:45 pm. How do I get home, and is the last train gone?"

MVP: Door-to-desk with the right entrance; cool corridors (shortest or coolest walk); heat now and next 24 h.
Later: Last-way-home guardian; heat-safe shift planner with Gujarati and Hindi voice alerts for outdoor staff.

Residents

small but growing; ~22% of planned built-up area

"A tower is going up next door. Will it take my sun?"

MVP: Coolest hour to go out; neighbourhood essentials with walking heat cost.
Later: My home's sun profile; future view from my window; report an issue on the 3D map.

Visitors

business guests, delegations, candidates, hotel guests

"Which gate, which tower, which lobby?"

MVP: Find the exact place, with a shareable meeting pin; a guided 3D story of GIFT City from 2016 to 2030.
Later: Arrival kit from airport and railway station; event mode; offline visitor card in three languages.

City operators

GIFTCL operations, C4, cooling, power, water, emergency services

"Heat wave tomorrow. Will district cooling hold?"

MVP: Heat-wave cockpit; tomorrow's cooling demand per building, labelled modelled.
Later, with partner data: Every C4 alarm in 3D context; restricted utility-tunnel view; calibrated cooling twin; evacuation planning.

Planners, investors, regulators

GIFTCL planning, developers, 939 registered entities, IFSCA

"What does a 30-storey tower on this plot do to its neighbours?"

MVP: Drop-a-tower what-if; growth vs cooling capacity by year; 4D city from 2016 to 2030.
Later: Plot feasibility with zoning; site selection by talent catchment; generative massing; permit pre-check.

One capability set for all

Shared: One time control from past to forecast to 2030; ask-the-twin agent that cites sources; provenance on every value; English, Gujarati and Hindi; open API.
Privacy: No accounts in the MVP. Routes are computed on the phone; home and office never leave the device (DPDP).

Which data each user needs Data to user matrix

Filled dot: needed for that user's MVP. Ring: needed later. Buildings with heights, footpaths with shade, weather and transit timetables unlock the MVP for four of the five groups, which is why they are the critical path.

Data	Workforce	Residents	Visitors	Operators	Planners
Buildings and heights	●	●	●	●	●
Entrances and drop-offs	●	○	●	○	○
Footpaths with shade	●	●	○	○	●
Weather forecast	●	●	○	●	●
Air quality	●	●	○	○	○
Transit timetables (GTFS)	●	○	●	○	●
Amenities	●	●	●		○
Imagery time series		○	○	○	●
Master plan and zoning		○	○	○	●
Cooling SCADA, C4 alarms, tunnel BIM		○		○	○

The deep-research report adds emergency services, tourists and city administration as separate stakeholders. We fold emergency services and administration into operators, and tourists into visitors, because they share data and interface; their extra needs (evacuation modelling, executive KPIs) sit on the operator roadmap.

What it does

One twin core, two faces, one agent. The core turns public data into a living model of GIFT City. The faces present it to two very different audiences, and the agent answers questions about it in plain language.

Product shape. Every face reads the same core through the same APIs. The partner arrow is dashed because the adapter exists before the data does: that adapter is the partnership pitch to GIFTCL.

Feature status

Built working in the proof of concept Feasible now open data is enough Needs field data we can collect it Needs partner GIFTCL or utility data

Feature	Status	Approach	Who it serves
Multi-source buildings with confidence	Built	Each source is a candidate; a ranked fusion picks a winner and keeps the rest with their evidence.	Everyone
Fresh imagery and 13 index layers	Built	Earth Engine composites, 1 m super-resolved scenes, swipe compare 2023 vs 2026.	Planners
Roads from 1 m imagery	Built	Our SegFormer model, mask to graph, every edge scored against OSM.	Planners
4D city, 2016 to today	Built	Yearly satellite-derived heights, calibrated; planned towers as ghosts once the master plan is available.	Planners, visitors
Metro, signals and traffic animation	Built	Timetable to motion profile to CZML; signals evaluated per frame. Signals and traffic are synthetic.	Operators, planners
Weather, India AQI, heat comfort (UTCI)	Built	City conditions card; UTCI on a 20 m grid using every tower's shadow.	Everyone
What-if: place a building	Built	Live shadow plus impacts: floor area, occupants, cooling vs capacity, cars, metro load, neighbours losing sun.	Planners, residents
Where to open a shop	Built	Multi-criteria scoring on a 60 m grid; rules written in plain English, validated before running.	Investors
Scenario Lab	Built	Baseline vs scenario with KPIs, series, narrative and assumptions. Results below.	Operators, planners
Shade-aware walking	Feasible now	Shade per footpath segment every 15 min; routing runs on the phone. Gated by spike SP-02.	Workforce, residents
Growth vs cooling capacity	Feasible now	EnergyPlus prototype buildings on Ahmedabad weather, scaled by floor area, summed with diversity.	Operators, planners
Ask the twin	Feasible now	MCP server exposing the same endpoints as tools; a 50-question known-answer test set.	Everyone
Door-to-desk and last-train alert	Needs field data	Hand-built GTFS from official PDFs plus OpenTripPlanner; entrances mapped on site.	Workforce, visitors
Real sensor fusion	Needs field data	Optimal interpolation already runs; needs a calibration walk with loggers.	Operators
Spatial C4 and calibrated cooling	Needs partner	NGSI-LD adapter into the same reading model; restricted deployment.	Operators
Utility-tunnel view	Needs partner	Restricted IFC/CAD; never public.	Operators

What the proof of concept already answers

Each Scenario Lab question runs a baseline and a scenario through the same engine. Grey is today; marigold is the scenario. Bars are drawn to scale within each card.

Transit

What if the metro ran every 15 minutes, all day?

Today153 min

Scenario7 min

Average wait for a train home. Working-day minutes with no train within 30 min fall from 78% to 0%; about 256 cars an hour leave the peak.

modelled Timetable reconstructed from news; mode shift is an elasticity estimate.

Traffic

Do coordinated signals cut waiting?

Today28 s

Scenario17.5 s

Signal wait per trip, down 37% with optimised offsets. On held-out trips: 29.4 to 20.2 s. A textbook green wave managed only 28 to 27 s.

synthetic Signal plans and demand are seeded; no queueing model yet.

Heat

How much cooler with shade trees?

Today93%

Scenario82%

Share of street cells at or above 35 °C at 14:00. Treated streets are 1.1 °C cooler in air temperature.

modelled Air temperature only; the larger radiant comfort gain is not yet modelled.

A finding that shapes the product: on 29 Oct 2026, only 3 to 7% of GIFT City's ground was in building shade at any daytime hour. Modelled comfort (UTCI) at 14:30 was 38 °C in sun and 34 °C in tower shade. Shade is scarce, which strengthens the case for trees and canopies, and is exactly why shade routing must pass its go/no-go test before it becomes the headline feature.

How it works

We borrow the shape of a proven open smart-city platform, Snap4City, and right-size it for one district and a small team. Data moves through six stages; each stage has one owner in the code and one rule it never breaks.

Same stages, lighter parts. Snap4City proves the six-stage shape at city scale across 40+ deployments. We keep the stages and the ontology-first idea, and swap a VM-heavy PHP and Java stack for a Python package, a file store that grows into PostGIS, and two web clients.

What we take from Snap4City

An ontology-first entity model (their Km4City; our NGSI-LD URNs) so every sensor, building and stop is linkable.
A broker-ready front door with an auth filter, so partner feeds arrive through one guarded entrance.
A 3D scene built from separate buildings, so each one can be picked, inspected and replaced in a what-if. Their note that fused photoreal meshes block building substitution is why we build our own base.
What-if as versioned scenarios keyed by time.
One-command deployment from a single machine upward.

What we leave

The reference footprint: an 8-CPU, 500 GB virtual machine and a PHP, Drupal, Apache and Tomcat stack. Too heavy for one district and one developer.
Visual Node-RED flows for core logic. We keep logic in tested code so automated tests can guard it.
AGPL coupling. Linking AGPL services would set our own licence before we have decided it (Q-10).
No-code dashboards as the main interface. Our users need answers and routes, not widgets.

Target containers for the MVP. The gateway is the one place that decides who sees what: public, internal or restricted. The Companion's routes never pass through it; they are computed on the phone from cached tiles.

Four flows that matter

Base data build, batch

Fetch each source; write raw files with a manifest (URL, date, SHA-256, licence).
Validate schema, coordinate system and geometry.
Fuse footprints and heights by trust rank; keep every candidate.
Write entities with stable URNs; build 3D Tiles, vector tiles, shade tables, routing graph.
Publish a versioned build (build ID, git SHA, manifests).

Shade-aware route, on the phone

Load the routing graph and this season's shade table once; cache for offline.
A Web Worker runs time-dependent A* with the user's shortest-to-coolest setting.
Show the route, % shaded and extra metres.
No trace is sent to any server.

Heat stress, hourly

Fetch the forecast; compute UTCI per street segment from now to +24 h.
Combine weather with each segment's shade and surface type.
Store in TimescaleDB; serve as tiles with method and run ID.

What-if, on request

The console posts a footprint, height and use.
A worker recomputes shadows, broken shade routes, cooling and trips.
Results are stored per scenario ID and diffed against the baseline build.
Target: answer within 10 s.

Rules the design never breaks

ADR	Decision	Why it matters	Status
0005	Provenance on every value	Modelled numbers mistaken for measured ones would destroy trust with a government audience.	Proposed
0006	Classification enforced server-side	Utility tunnel, substations and CCTV are critical infrastructure. Hiding a layer in the UI is not security.	Proposed
0007	Build our own 3D base	Google's photoreal tiles do not cover India, and their terms forbid measuring from them.	Proposed
0008	NGSI-LD shaped live entities	BIS IS 18003 and IUDX are NGSI-LD based; this is how we plug into Indian government platforms.	Proposed
0010	Simulation: plans in, state out	Deterministic, seeded engines make every scenario reproducible and testable.	Accepted
0011	LLM proposes, schema decides, human reviews	The model never computes a number we show; it writes rules that are validated and editable.	Accepted
0013	Air quality is city-wide; comfort is spatial	One station 4.5 km away cannot draw a street-level AQI map; heat comfort can.	Accepted

Data model, identity and security zones Engineering

Core entities: Building, Storey, Entrance, FootpathSegment, RoadSegment, PublicSpace, TransitStop, TransitRoute, Amenity, WeatherForecast, AirQualityObserved, ShadeSample, HeatStressForecast, Scenario, DataSource. Every record carries the same envelope:

{
  "id": "urn:ngsi-ld:Building:giftcity:B-GIFT-ONE",   // stable forever, never derived from geometry
  "type": "Building",
  "geometry": "EPSG:4326 stored; EPSG:32643 (UTM 43N) for analysis",
  "valid_from": "2013-01-01", "valid_to": null,       // 4D: history, today, planned
  "status": "operational | under_construction | planned",
  "provenance": { "kind": "measured | derived | modelled | synthetic",
                  "source_ids": ["S24", "D41"], "method": "registry", "run_id": "…" },
  "classification": "public | internal | restricted",
  "xref": { "osm": "way/…", "open_buildings": "…" }   // external IDs are cross-references only
}

Zone	Contents	Access
Public	Companion, public tiles, public API	Anonymous, rate-limited
Internal	Console, scenarios, agent with internal tools	Signed in (Keycloak, OIDC)
Restricted	Partner data: utilities, C4, tunnel BIM	Named users, audited, separate storage, India region

Standards mapping: buildings as CityGML 3.0 / CityJSON and NGSI-LD Building; footpaths as CityGML TrafficSpace and NGSI-LD RoadSegment; weather and air as Smart Data Models WeatherForecast and AirQualityObserved; transit as GTFS; sensors as OGC SensorThings; building systems later as Brick and IFC.

As-built proof of concept and its API Engineering

Layer	Code	Rule
Sources	gifttwin/sources/*	One module per source; every download gets a manifest.
Domain	gifttwin/buildings/*	All height candidates kept; the winner is explained.
World	gifttwin/world.py	The only runtime file reader; swap for PostGIS here and nothing else changes.
Simulation	gifttwin/sim/*	Deterministic and seeded; CZML out.
Analytics	gifttwin/analysis/*	Every number carries assumptions and provenance.
Integrations	gifttwin/integrations/*	External ML runs behind a process boundary in its own environment.
Client	frontend/src/*	React never touches Cesium directly; it updates a store or calls a controller.

Endpoint	Purpose
GET /api/layers/{buildings,change}	Fused buildings and change polygons as GeoJSON
GET /api/world/{context,imagery,roads}	Boundary, metro, roads, POIs; imagery catalogue; road-extraction stages
GET /api/sim/{transit,vehicles}.czml	Time-dynamic metro and traffic
GET /api/sim/{sensors,heat,conditions,comfort}	Readings, fused heat field, weather and NAQI, UTCI raster
GET /api/scenarios	Scenario Lab results with assumptions
POST /api/whatif/building	Impacts of a proposed building
POST /api/siting/evaluate · /api/rules/translate	Site suitability; plain English to validated rules

Path to target without a rewrite: PostGIS and TimescaleDB when scenarios are saved or layers pass 50 MB; 3D Tiles past 10,000 buildings or for façade shaders; an NGSI-LD adapter at the first partner feed; SUMO when real traffic counts exist; the MapLibre Companion at the first public pilot.

Data and research

Two days of research and build produced ten research notes, a 55-entry data catalogue and 99 cited sources. Every fact is tagged by how sure we are, and contradictions are kept in the open rather than smoothed over.

        [V] verified at source[R] reported by others[I] our inference[U] unverified[K] background knowledge
      

What we collected

Dataset	What we hold	Used for	Provenance
OpenStreetMap	Pulled 28 Oct 2026: boundary polygon (3.68 km²), a 7,630-node road graph, metro viaduct, POIs, 61 buildings inside the boundary.	Network, context	measured
Google Open Buildings v3	10,580 footprints in the box, 1,661 inside the boundary, streamed from an 8.2 GB shard in about 12 s.	Footprints	derived
Open Buildings 2.5D Temporal	Yearly heights 2016 to 2023 per footprint via Earth Engine; calibrated ×1.24 against 9 trusted towers.	Heights, 4D	modelled
Curated registry	43 buildings with floors, height, status and web sources.	Authoritative heights	measured
Sentinel-2 via Earth Engine	Oct 2026 composite; single scenes 1 Jun 2023 and 15 Oct 2026; 13 index layers (NDVI, EVI, NDBI, BSI, NDWI, MNDWI, land surface temperature, and 2023→2026 changes).	Greenery, heat, change	derived
S2DR3 at 1 m	Two super-resolved scenes (2023, 2026), true colour plus 12 bands; 53 of 75 monthly credits left.	Roads, future 3D	modelled
Roads from our model	Route Resilience SegFormer on 1 m scenes: 92–95% of detections lie on OSM roads; finds 48–51% of OSM.	OSM edit checklist	modelled
Change 2023 → 2026	Spectral change polygons; 184 buildings flagged as changed.	Construction radar	derived
Weather and air	Open-Meteo hourly cache; CAMS pollutants with India NAQI computed by us: 120 Moderate, ozone-led, at 14:00 on 29 Oct while US AQI read 73.	Conditions, UTCI	modelled
Basemap	Esri World Imagery over GIFT City is Vantor 0.34 m from 16 May 2026 (non-commercial demo terms).	Visual reference	measured
Transit timetables	Four official PDFs (GMRC, GIFT Go, GSRTC, intra-city bus) located; GTFS to be hand-built.	Door-to-desk	to build
Master plan, cooling SCADA, C4, drone survey	Not available. Routes: RTI for the plan; partnership for live feeds; Digital Sky check for drones.	Zoning, operators	blocked

How a building gets its height. No source is thrown away. The fused record says which source won and why, so a planner can check it and a field survey can overrule it.

The weak spot, shown plainly

High 4 Medium 7 Low 531 None 1,176, mostly sheds

Shadows, heat comfort and what-if are only as good as these heights. The roughly 60 main towers are few enough to verify by hand, which is the first field task on the roadmap.

Key research facts we build on Research

Topic	Fact	Tag
Scale	~25,000–30,000 workers; 25 operational buildings, 37 under construction; ~62 M sq ft planned, ~29.5 M allotted; use mix ~67% commercial, 22% residential.	[R] S04 S10
Operations brain	C4 command centre integrates 70,000+ I/O points across power, cooling, water, waste, utility tunnel, streetlights and GIS.	[V] S82
District cooling	Planned 180,000 TR; phase 1 since 2015; claims electrical demand cut from 240 to 135 MW; SCADA metered.	[V] S06
Utility tunnel	~16 km (World Bank) or ~17 km; power, water, sewage, waste pipes, chilled water.	[R] S08
Metro	GIFT City branch opened Sep 2024; peak-only every 49–57 min in Jan 2026; operation until 11 pm approved from May 2026, branch details unknown; Shahpur extension approved Feb 2026.	[R] S12 S13 S15
Heat	Ahmedabad's Heat Action Plan (2013) followed the May 2010 heat wave at 46.8 °C; it gives us alert thresholds and credible partners.	[V] S52
Shade routing evidence	CoolWalks (Sci. Reports, 2025): uniform grids gain nothing from shade routing; mixed heights can. GIFT City is mixed, so gains are plausible but unproven.	[V] S51
Maps law	DST 2021: data finer than 1 m horizontal or 3 m vertical must be Indian-owned and stored in India.	[R] S46
Personal data	DPDP Rules notified Nov 2025; main obligations from about May 2027.	[R] S50
Drones	GIFT City sits about 12 km from the airport, at the edge of the yellow band; assume permission is needed.	[R] S49 S85

How we reconciled the deep-research report Research

The external deep-research report is kept unchanged in docs/inputs/2026-10-29-deep-research-report.md. Each claim was checked against our research before use.

Adopted

Stakeholder to data to feature mapping; adds emergency services and city administration to our users.
Data-quality KPIs: freshness, completeness, latency, reliability, shown on a data-health page.
Standards stack: CityGML/CityJSON, IFC, SensorThings, NGSI-LD, 3D Tiles, OGC API. It matches ours.
Governance: catalogue, audit logs, a steering group once partners join.
Platform comparison, used in section 06.

Deferred, by trigger

Kafka, Kubernetes and a service per function: no event volume or team size justifies them yet.
Graph database: PostGIS plus URN relations cover our links.
Closed-loop control of city systems: a non-goal until an operator asks for supervised advice.
Blockchain provenance, AR/VR: no user need found.

Not used, unverified

"5% of assets yields 80% of benefit"; "payback in 15–18 months"; "28% water OpEx cut": no primary source.
Snap4City licence given as GPL: the repository is AGPL-3.0.
"NGSI-LD backbone": Snap4City is FIWARE-certified; the NGSI version is not stated.
Roadmap dates starting Jan 2025 are out of date.

Technology

Open source end to end, chosen for a geospatial and simulation workload, and sized for one district rather than a national platform. Heavier infrastructure arrives only when a named trigger fires.

Concern	Built today	MVP target	Why this and not the alternative
3D console	CesiumJS 1.146	Same, with 3D Tiles	Globe-accurate, real sun and shadows, time-dynamic CZML, Apache-2.0. Unreal needs GPU streaming and cannot reach phones.
Phone companion	Not yet	MapLibre + deck.gl PWA	Cesium is too heavy on a mid-range Android in the sun. One data core, two renderers (ADR-0003).
Web app	React 19, Vite 8, TypeScript strict, zustand	Same	Fast iteration; the store keeps React away from the 3D engine.
Backend	Python 3.11, FastAPI, pydantic, Typer CLI	Same	Geo, ML and simulation libraries are Python-native: shapely, pyproj, networkx, rasterio, pvlib.
Storage	Files with manifests	PostgreSQL + PostGIS + TimescaleDB; object storage	One database for geometry and time series. Orion-LD only when a partner feed is real.
Pipelines	CLI commands	Dagster	Asset-based pipelines with lineage; a scheduler only when there is something to schedule.
Remote sensing	Earth Engine; Gamma Earth S2DR3	Same, plus STAC catalogue	No raster downloads for composites; 1 m scenes at a few credits each.
Machine learning	Our SegFormer roads model in its own environment	+ ChangeFormer for new buildings	Process boundary: the twin never imports torch.
Journeys	Not yet	OpenTripPlanner 2	Mature multimodal GTFS routing. Shade routing stays on the phone.
Simulation	Own engines; optimal interpolation for sensors	+ EnergyPlus prototypes, SUMO later	Physics where it decides the answer; seeded and reproducible everywhere.
AI	Groq, gpt-oss-120b, JSON rules	MCP server + eval set	Swappable model; every number comes from a tool, never from the model.
Quality	pytest, vitest, Playwright; ruff; CI on GitHub	+ performance budgets in CI	86 automated tests today (66 backend, 13 frontend unit, 7 end-to-end), including Python and TypeScript parity for signals.
Platform	Local	Docker Compose, one India-region server, Keycloak, OpenTelemetry	India hosting for fine data and any personal data. Static Companion survives API downtime.

Build or buy

Platform	Strength	Why it is not our core	What we borrow
Snap4City	Open, FIWARE-certified, full city pipeline, 40+ deployments	Heavy VM stack; AGPL; dashboard-first	Six-stage shape, ontology-first model, scenario versions
FIWARE Orion-LD	The open NGSI-LD context broker; IUDX compatible	Plumbing only, no 3D or analytics	Our live-entity API shape; adopt the broker at first partner feed
Azure Digital Twins, AWS TwinMaker	Managed scale, enterprise support	Lock-in, cost, managed twin products get retired	Graph-of-entities thinking
Esri ArcGIS Urban	Best planning GIS; common in Indian government	Licence cost; weak on live data and citizen apps	Planning scenario UX
Bentley iTwin	BIM and infrastructure depth; owns Cesium	Cost; engineering-first UI	Restricted utility-tunnel view later

Scale triggers: when heavier infrastructure earns its place Engineering

Add	Only when
PostGIS + TimescaleDB	Scenarios are saved, time series accumulate, or a layer passes 50 MB
3D Tiles pipeline (pg2b3dm)	More than 10,000 buildings, or façade shaders such as lit windows
Orion-LD broker, MQTT	The first partner or field sensor feed is live
SUMO traffic	Real counts and signal plans exist to calibrate it
Kafka	Sustained event volume beyond what Postgres and MQTT handle, roughly thousands per second
Kubernetes	More than one team operates the platform, or uptime commitments exceed one server with backups
Gaussian-splat hero zone	Drone or phone capture is permitted; Cesium has supported splat tilesets with LOD since Apr 2026

Feasibility

The technology is proven by the running prototype. The open risks are data access, user evidence and positioning, and each has a planned answer.

Technical

High

Every MVP engine has a working prototype: 3D, fusion, shadows, comfort, simulation, what-if, LLM rules.

Data

Medium

Open data covers the MVP for four of five users. Heights are weak; master plan and live feeds are blocked.

Legal

Medium

ODbL share-alike, non-commercial API terms, DPDP, drone permission, critical-infrastructure care. All have rules in place.

Cost

High

Free tiers and one server run the MVP. Commercial use needs paid weather and imagery; the cost model is still open.

Adoption

Low to medium

No primary user research yet, and an official twin is coming. Interviews and a partner-first pitch close this.

Team and time

Open

Capacity, deadline and budget are undecided, so the roadmap below is indicative.

Top risks and what we do about them

Risk	Odds	Impact	Mitigation
Official twin launches overlapping public features	Medium	High	Lead with prediction and daily utility; offer integration through standards rather than compete on 3D.
Shade routing shows little gain	Medium	Medium	Spike SP-02 first: under 10% more shade at 15% extra distance demotes it to a layer; lead with heat stress and door-to-desk.
Wrong heights give wrong shadows	Medium	High	Curated registry, on-site floor counts for ~60 towers, golden tests.
Modelled numbers taken as real	Medium	High	Provenance chip on every value; "modelled" in the copy; hatching for synthetic layers.
Sensitive infrastructure exposed	Low	High	Server-side classification with tests; review before any layer is published.
Timetables change and mislead	High	Medium	"Timetable as of" dates; monthly re-check; feedback button.
Licence breach (ODbL, non-commercial terms)	Medium	High	Licence matrix; CI check that every source has a licence entry.
Scope creep	High	Medium	Slice gates, ADRs and a demo script written before the build.

Limits of the prototype, stated up front Honesty

Signals, vehicles, most sensors, shop POIs and rents are synthetic. They are labelled in the interface and deterministic by seed.
1,176 small footprints still use a placeholder height and are hidden by default.
Satellite-derived heights are calibrated on only 9 tower pairs, which is why they rank as low confidence.
The change layer is a 10 m spectral heuristic: good for "look here", not for counting buildings.
Esri imagery and the free Open-Meteo API are used under non-commercial terms.

Roadmap

Ruthless MVP, ambitious roadmap. Each slice ships on its own and must pass a measurable gate before the next one depends on it.

28–29 Oct 2026

Proof of concept

3D twin, fusion, imagery, metro, Scenario Lab, what-if, siting, conditions and comfort.

About 12 weeks

MVP

Slices S0 to S7 on open data and our own field data. Public Companion pilot; console demo for GIFTCL.

After MVP

v1, field data

Entrances mapped, sensor calibration walk, verified GTFS, last-way-home guardian, heat-safe shift planner.

With a partner

v2, live city

C4 alarms in 3D, calibrated cooling twin, restricted tunnel view, emergency planning.

Slice and its gate

W10

W11

W12

Indicative, assuming one or two developers. It will be fixed once team, time and budget (Q-03 to Q-05) are answered.

How we will validate before we claim Validation

What	How	Pass mark
Is the problem real?	10–15 interviews across levels including service staff; a week of commute diaries; observation at 09:00, 13:00, 18:30, 20:00	Top three pains confirmed or replaced
Shadows	Timestamped photos at 5–10 points vs predicted shadows	Edges within a few metres
Heat comfort	Walk with a temperature and humidity logger plus a black-globe thermometer in sun and shade	Error reported in the docs
Journeys	20 timed trips vs the plan	95% within 5 min
Ask the twin	~50 known-answer questions on every change	Every number traced to a tool call
Performance	Real mid-range Android outdoors; budgets in CI	3 s to interactive; 300 KB initial JS

Decisions needed

Engineering can move ahead on the foundations today. These choices decide what the MVP looks like, so they come before slice S1.

Q-01

What is the endgame: portfolio, competition, startup or a government pitch?

Recommendation: treat it as a government and GIFTCL pitch with a public pilot. It sets the bar for honesty and security that every other endgame also needs.

Q-02

Who is the primary user?

Recommendation: the commuting professional, including service staff. Largest group, sharpest documented pain, and served by open data alone. Planners stay the second face.

Q-03 to 05

Team, time and budget

Recommendation: one lead developer plus one part-time field and research helper for 12 weeks; budget for one India-region server, paid weather if commercial, and a few S2DR3 credits.

Q-12

When to approach GIFTCL, DEV IT and GMRC

Recommendation: after S2, with a working Companion and the console demo. Lead with the NGSI-LD adapter as the integration offer.

Q-09, Q-10

Product name and our own licence

Recommendation: an independent name that cannot be mistaken for official branding; Apache-2.0 for code, ODbL where outputs derive from OSM.

The short version