CGIAR Agricultural Risk Intelligence Tool

The problem

CGIAR generates and consumes massive volumes of unstructured agricultural and climate data — bulletins, field reports, sensor series, papers, programme metrics — which teams used to process by hand to produce actionable risk reports. The cycle was long, fragile, and bottlenecked on one or two senior analysts.

The approach

I designed and deliver a serverless NestJS + Next.js monorepo on AWS (Lambda, API Gateway, Bedrock). The core is a RAG pipeline over climate and agronomy sources, with Claude-powered agents that extract, normalize, and surface risk in a format senior stakeholders can act on. The application layer runs on Anthropic Claude and AWS Bedrock, orchestrated in TypeScript and Python.

• Continuous ingestion and normalization of heterogeneous sources.
• Typed RAG with an AWS-hosted vector store, evaluated against internal ground truth.
• Agents with tool use that invoke internal APIs, query the vector store, and generate artifacts.
• Structured outputs validated against schemas — no free-text drift.
• End-to-end observability: trace, latency, cost per invocation.

The outcome

• What used to take hours or days of manual analysis now runs in minutes of automated pipeline.
• Stakeholder-ready outputs without a human in the drafting loop — the agent ships the final artifact.
• A reusable stack for future internal CGIAR tooling; the RAG and agent components are now a building block.
• I defined the team's engineering standards (Claude Code workflows, automated quality hooks, monorepo conventions) and mentor engineers across multiple concurrent initiatives.