Parvind Dutta

Investigating a counterparty’s real structure—SPVs, interlocking directorates, regulatory history—means reconciling scattered filings by hand, with no way to tell a confirmed fact from an assumption.

This platform renders that structure as a navigable graph, where each fact carries an explicit evidence grade (primary-source, independent verification, first-party disclosure, or analytical hypothesis) instead of uniform certainty. An assistant answers questions grounded in that graph, not unguided model output. Open gaps are tracked as structured items—what’s missing, what it would settle to, how hard it is to obtain—rather than left as free text.

Back end: Python/FastAPI on Azure Kubernetes Service, Azure OpenAI, Entra ID.
Front end: React/TypeScript PWA. A native iOS client is in development, sharing one schema end to end. A hypothesis is never automatically upgraded to fact; sensitive material is scoped to authorized viewers and enforced server-side.

SagePAS agent portal enables appointed agents to quote, bind, issue, take payments, and service policies in a single workflow. The platform connects rating, policy issuance, and payment capabilities while maintaining a shared policy record with the policyholder’s retail experience, eliminating the lag and reconciliation problems created by disconnected systems. The platform is live in production and supports real insurance transactions today. My role spans product strategy, technology architecture, solution design, and delivery leadership, bringing together insurance-domain requirements with modern cloud-native engineering to create an integrated agent and policyholder experience.

Built SageFNOL, the first notice-of-loss processing engine to streamline claims intake from document submission through extraction, validation, and CRM handoff. The pipeline uses Azure Content Understanding for document extraction with a GPT-4 Vision fallback for image-heavy submissions, ensuring reliable processing across photos, scans, and PDFs. Every extracted field is traceable to its exact source location in the original document, and every knowledge-grounded conclusion is tied to the supporting passage for full auditability. The solution was built in Rust on a shared PostgreSQL schema with JSONB stage storage, Azure Blob for document storage, and a durable transactional outbox pattern to sync finalized claim data into HubSpot as claim tickets.

Built a document-driven underwriting pipeline covering classification, extraction, detection, analysis, scoring, and action as 6 independently inspectable stages. The system uses Azure Content Understanding with a GPT-4 Vision fallback to process image-only submissions and applies knowledge-assisted scoring to ground debit and credit factors in supporting underwriting material. Every output remains queryable by stage, enabling underwriters to inspect how a recommendation was formed and trace conclusions back to both source document regions and supporting knowledge. The solution also introduced a hand-authored commercial property submission schema aligned to real underwriting workflows and syncs final decisions into HubSpot as deal records through a durable outbox architecture.

Built the operations command center behind MapleSage’s flagship insurance platform, a unified B2B workspace designed to give insurance teams one real-time view across the policy lifecycle. Rather than stitching together separate systems, the platform is built around a shared record model that connects quoting, claims intake, underwriting, policy servicing, payments, and CRM visibility in one operational surface. It supports live intake queues, funnel and service monitoring, and cross-functional workflows that let teams act from a single workspace instead of switching between disconnected tools. The platform is part of MapleSage’s broader architecture in which claims and submissions are created as native HubSpot objects while the insurance machinery runs underneath, creating a tightly connected operating environment for service, underwriting, and distribution teams.