Hazard Forecasting & Property Intelligence Platform
Our client provides hazard forecasting and property intelligence that insurance companies use to assess and price risk. We built their web application end to end — the platform used to pitch prospective customers on how an AI agent can estimate damage from natural calamities in real time.
The agent draws on underlying data from Google, DeepMind, and Google Earth Engine, and orchestrates the ML team's suite of damage-prediction models to turn a user's question into a concrete, defensible risk estimate. Because these workflows drive real coverage and pricing decisions, reliability, security, and data trustworthiness map directly onto our client's ability to win and keep customers.
Architecture overhaul
As the architect on this engagement, I owned the technical direction of a platform-wide architecture overhaul. We migrated the API layer to a modern backend framework that gives the LLM agents and the user interface a single shared contract — so the two stay perfectly synchronized as the product evolves — and introduced a new access-control layer robust enough for sensitive property and risk data while staying invisible to end users.
Agentic layer & dynamic frontend
A core piece of the work was the agentic layer and the frontend that sits on top of it. Building on Google's Agent Development Kit (ADK), we wired the underlying prediction models and geological/hazard data sources directly into the agent for smooth, fast agentic AI interaction, then built a dynamic frontend that renders components on the fly from the agent's outputs — turning a raw response into a detailed, interactive report the client's underwriters can explore rather than just read. We also made the agents and the surrounding system scalable, and ran stress tests to confirm the experience stayed smooth for end users under load.
Operational visibility & quality
Alongside this, we invested heavily in operational visibility and quality. We instrumented telemetry-based audit logging so the team could rapidly diagnose system inefficiencies and API failures, cutting the time to identify and resolve production issues before they reached end users. We rebuilt user-journey tracking and retention analysis with full PII redaction, giving the client clear, privacy-safe visibility into how underwriting teams actually use the platform — insight that now feeds directly into product and roadmap decisions.
We aligned the interface with the shared design vision for a more polished, consistent experience that reduces friction for professionals working with dense risk data, raised unit test coverage from roughly 40% to 90% to materially reduce regression risk and increase confidence in every release, and ran rigorous evaluation testing on the LLM agent to confirm production readiness — ensuring the automated damage estimates and generated reports meet the accuracy bar an insurance decision-making context demands.