How I’d build it
A platform a bank can put in front of its validators.
Climate analytics deals are rarely lost on the map. They are lost when a model-risk team asks where a number came from and the answer is “the vendor’s model”. So I design for explainability, lineage, residency and validation evidence first — with the science team’s methods turned into tested, versioned code.
Why clients buy now
- SingaporeISSB-based climate reporting through SGX and ACRA — Straits Times Index issuers from FY2025, other listed issuers phased to FY2030, limited assurance on Scope 1–2 from FY2029.
- AustraliaAASB S2 mandatory for Group 1 entities from 1 Jan 2025; Groups 2 and 3 follow in 2026 and 2027.
- SupervisorsMAS environmental-risk and transition-planning guidelines (effective Sept 2027) and APRA CPG 229 expect the analysis inside credit and underwriting, not just the report.
- NatureThe ISSB published its nature-related exposure draft in Oct 2026, building on TNFD. Location-level evidence becomes a reporting input.
Once assurance arrives, “disclosure-grade” means an auditor can trace every reported number to its data, method and run. That is an engineering requirement.
Architecture
Eight layers, one region, one lineage.
Open and licensed data are pre-processed where they live and reduced into Singapore. Everything a client sends stays there. Every layer writes to the same provenance record.
- Data foundations
- Engines — risk, nature, AI
- Delivery to clients
- 01IngestManifests, checksums, licence IDs, STAC catalogue of every layer and version
- 02StoreCOG rasters · versioned Zarr cubes · GeoParquet / Iceberg results
- 03AssetsUpload → validate → geocode with a precision tier → tenant-scoped registry
- 04ComputeOrchestrated batch sampling at scale; bounded on-demand scoring
- 05Risk engineHazard × exposure × vulnerability → losses, return periods, scenarios
- 06NatureTNFD LEAP Locate and Evaluate · deforestation and EUDR checks
- 07AIEarth-observation models · RAG over standards and methodology · tool-calling agents
- 08ServeREST API · map tiles · bulk GeoParquet into the client’s lake · disclosure packs
Principles
Six rules I would not trade away.
01
Lineage from day one
Every output row carries a run id, methodology version and input snapshot ids. Retrofitting provenance is a rewrite — and it is the first thing a model validator asks for.
02
Losses before scores
Compute hazard intensity, damage and loss first; any band or score is a display layer. A single opaque score is the fastest way to lose a credit committee.
03
The LLM never owns a number
Language models explain, retrieve and draft. Every figure comes from a typed tool call against the platform’s own results, with a citation.
04
Data gravity one way
Pre-process open data next to where it lives, copy reduced APAC subsets into Singapore, never move client locations out of region.
05
Licence manifest per output
Every derived layer knows which upstream licences it inherits. Several “open” datasets are non-commercial; finding out after launch is expensive.
06
Open formats, not managed features
COG, Zarr, GeoParquet and STAC on object storage. Managed geospatial services get retired; open formats do not.
The first 90 days
Listen, ship, then make it bank-ready.
- 1
Days 1–30
Listen, inventory, lay foundations
- Learn the science team’s methods and what has been promised to clients.
- Inventory stack, data licences and partner contracts; flag anything that blocks selling to banks.
- Infrastructure as code, CI, storage layout, and a provenance table every job writes to.
- A “golden portfolio” whose outputs are diffed on every change.
- Start SOC 2 readiness — it gates enterprise deals and takes months.
- 2
Days 31–60
Ship the first honest increment
- Flood, heat and water stress for Singapore plus one market, end to end.
- Portfolio upload with location-precision tiers; per-asset results and losses.
- Methodology v0.1 written with the scientists, versioned with the code.
- One design-partner institution running its own sample.
- 3
Days 61–90
Make it pilot-ready
- REST API, bulk export, and a validation pack a model-risk team can work through.
- RAG over TNFD, ISSB and our methodology — with citations and evaluations.
- First two or three hires made; partner scope written as outcomes, not hours.
- A board-level view of budget, risks and the 18-month roadmap.
18-month roadmap
Each phase ends with something a client can verify.
See how the technology evolves, startup to enterprise →Foundations
IaC, CI, storage, orchestration, provenance, golden portfolio, SOC 2 readiness
Exit: A run produces a provenance record and a diffable output
MVP
Flood, heat, water for two markets; portfolio upload; map UI; methodology v0.1; TNFD Locate
Exit: Two design partners can explain any number to the asset
Pilot-ready
API; validation pack; nature Evaluate and EUDR; RAG with citations; SOC 2 Type I → II; pen test
Exit: A paid pilot clears vendor due diligence
Scale
Licensed flood and cyclone models; millions of assets; insurer event sets; Sydney stack; disclosure copilot; ISO 27001
Exit: One institution in production; restatement process proven
First hires, in order
One scientist and one ex-bank modeller change sales outcomes.
- 1Geospatial data engineerOwns ingest, formats and sampling at scale
- 2Climate / hazard scientistDefends methodology to bank validators
- 3Risk quant, ex-bank or insurerTurns hazard into loss, PD/LGD and capital language
- 4Full-stack engineerMap UI, portfolio workflows, review queues
- 5Platform engineerTenancy, identity, CI/CD, observability, cost
- 6ML / LLM engineerEarth-observation models, retrieval, evaluation harnesses
External partners
I have been the partner. I scope work as testable outcomes.
Decisions I would make early
- Python vs TypeScript
- Python for pixels, geometry, risk maths and LLM pipelines; TypeScript for the web app; a generated API client is the contract.
- Operational vs analytical store
- PostGIS for assets, tenancy and tiles; GeoParquet / Iceberg for results history and bulk delivery.
- Orchestration
- Asset-based (Dagster) so every data product has lineage and per-partition status.
- Earth Engine vs self-hosted
- Self-hosted open pipeline in production for residency and reproducible reruns; Earth Engine for exploration.
- Hazard data
- Open layers labelled “screening grade” for the MVP; license a commercial flood model before credit or underwriting use.
- LLM placement
- In-region model access behind an internal gateway with logging and redaction; no browser-to-model calls.
What banks and insurers will ask for
- MAS Technology Risk Management and Outsourcing Guidelines; APRA CPS 234 and CPS 230
- SOC 2 Type II, then ISO/IEC 27001; independent annual penetration tests
- Model documentation and validation support aligned with the MAS AI Risk Management Guidelines (7 Oct 2026)
- Data residency in Singapore and Sydney; per-tenant keys; immutable audit log
- A documented methodology-change and restatement process
- A licence manifest proving every delivered layer may be used commercially