ConservAxion — verifying clean-energy and biodiversity impact with Azure AI Foundry

Pilot status: in active deployment in KwaZulu-Natal, South Africa. Specific cohort numbers and field results are deliberately not in this post. The pilot is small, the figures move weekly, and a snapshot now would mislead. A separate piece will follow when the cohort closes.

The trust problem#

Conservation credits and clean-energy attestations are two markets held together by self-reported PDFs. A donor writes a cheque. A field team does the work. Months pass. A report arrives. Most of the time the report is true. Sometimes it isn’t, and the market quietly absorbs the cost.

ConservAxion is a pilot platform, built with Craig Beech of CBIO, that takes a different angle. Instead of better reports, it produces better evidence: signals captured from the field, validated by AI, anchored in a tamper-evident audit trail. A donor does not have to trust the field team’s prose. They can verify the chain itself.

The shape#

The platform is built on Microsoft Azure end-to-end. Azure Static Web Apps for the Entra-protected donor surface. Azure Functions for the back end. Cosmos DB for state. IoT Hub for device telemetry. Azure AI Foundry, via Azure OpenAI, for validation. Azure Confidential Ledger for the tamper-evident audit trail. Key Vault for the one secret in the whole architecture that cannot be replaced by a managed identity.

The architectural moves that turned out to matter:

Multiple independent evidence streams. Solar telemetry from pilot inverters. Custodian-submitted field photos. Periodic survey data. Remote sensing. No single stream can mint a credit on its own. They meet at a compliance engine that produces a single verdict — pass, review, or reject — and downgrades to “review” the moment any stream goes silent.

AI as judge, not author. Foundry agents are only ever called to validate. They score telemetry plausibility against recent history. They compare before-and-after photos and emit a change verdict. They classify whether a survey image plausibly shows the claimed location. Every call returns a confidence score. A tunable threshold decides whether the outcome auto-approves, sits for human review, or is rejected. The AI never produces text that becomes the record.

Tamper-evident audit trail. Every consequential decision is hashed and written to two stores. One durable, one append-only and tamper-evident. An auditor twelve months from now does not have to take our word that the chain hasn’t been edited. The platform guarantees it.

Managed identity throughout. One long-lived secret, scoped to a single function whose only job is to talk to a third-party inverter cloud that cannot federate. Everything else is system-assigned managed identity, scoped to the smallest blast radius that works.

Why this shape#

The Foundry choice matters because the donors and impact-finance institutions that ConservAxion talks to do not want to debate model behaviour. They want the AI to be constrained, its job to be bounded, and the chain of evidence to not pivot on whether a model hallucinated. AI-as-judge is what makes that conversation possible.

The Confidential Ledger choice matters for the same reason. A donor auditing a credit later does not need to trust that the proof chain has not been edited. The platform guarantees it.

The rest of the stack is plain Azure architecture, picked so a small team can keep it running without watching it like a hawk.

The pilot#

The platform is small on purpose. One partner organisation, a handful of inverters in pilot households in KwaZulu-Natal, a simulator that exercises the same pipeline. Pilots that try to demonstrate impact at scale tend to demonstrate scale and miss the point. This pilot is the minimum shape that lets us test the claim at the centre of the platform: several independent evidence streams, joined by a compliance engine, and anchored in a tamper-evident ledger, producing a credit whose provenance an auditor can follow.

It has not been quick to get right. There is a separate set of notes, somewhere, that I will probably never publish, for the small platform quirks every Azure project meets eventually.

Closing#

ConservAxion is the production work behind a year of writing, speaking, and community organising in Durban around Microsoft Foundry. Craig Beech of CBIO has provided the corroborating endorsement letter attached to the MVP submission this article is part of.

It is still pilot-stage. The questions that matter most — the cohort, the funder reception, the field results — do not have public answers yet. When they do, there will be a separate piece. Until then, the architecture above is what is on the table.

Durban. With thanks to Craig Beech of CBIO for the field partnership, and to the Microsoft MVP programme for the Azure credits that keep the pilot running.