AI Search Platform
Multi-Source Agentic Research Engine
Enterprised Perplexica (open-source AI search) for internal use by adding agentic query classification, multi-source routing, streaming cited answers with contextual widgets, and an MCP server for direct AI assistant integration. Deployed across 3 AWS accounts with hub-and-spoke Terraform.
Internal teams needed a unified search tool that could surface information from multiple source types with rich context, not just keyword matching. No existing solution combined query classification, multi-source routing, and contextual widgets.
Lead engineer. Full-stack app, agentic search pipeline, model provider abstraction, MCP server, and hub-and-spoke Terraform across 3 AWS accounts.
Highlights
Agentic query classification
LLM classifies each query to determine the best search sources (web, academic, discussions) and which contextual widgets to display. Generates standalone reformulations for better retrieval quality
Real-time streaming with citations
Responses stream in chunks with source attribution and provenance tracking. Intermediate reasoning blocks show the agent's search process transparently
Contextual widget system
Weather, interactive stock charts, calculations, and curated news inject automatically based on query classification, driven by classification rather than keyword matching
Dual API exposure
REST endpoints for the web interface plus an MCP server for direct AI assistant integration. API key management with usage analytics and daily stats
Reusable SSO Terraform module
Built a 13-task Entra ID Terraform module from scratch: app registration, Graph API permissions, RBAC groups, Secrets Manager integration, feature-flagged across DEV/STG/PRD. Hub-and-spoke CI/CD account assumes deployment roles into target accounts