How a solo founder runs a five-continent tender platform on AlloyDB and MCP
A solo-founder startup reduced query latency 47-fold and automated operations by consolidating global tender data on AlloyDB with MCP, eliminating separate databases and dedicated staff.
Lucius AI, a tender-intelligence platform operating across five continents, consolidated its relational catalog, audit logs, and vector embeddings into a single AlloyDB for PostgreSQL instance. The move eliminated the need for separate relational, vector, and log databases, simplifying infrastructure while maintaining strict Cloud IAM authentication. By centralizing data in AlloyDB, Lucius AI established unified backup schedules and identity management, reducing operational overhead for a solo founder.
Semantic search performance improved dramatically after migrating to a ScaNN index in AlloyDB, cutting query latency from 1.14 seconds to 24 milliseconds—a 47x speedup. The AI agent, connected via Model Context Protocol (MCP), automated index recommendations and performance audits, further streamlining maintenance. Embeddings are refreshed weekly using AlloyDB AI’s ai.refresh_embeddings, while columnar engine acceleration automatically identified and stored 40 frequently queried columns in memory within a day.
The AI agent, configured with the MCP Toolbox for Databases, handles routine tasks such as data freshness checks, incident forensics, and on-demand analytics via ad hoc SQL queries. Access controls restrict the agent to SELECT permissions across the schema and UPDATE on a single operational table, preventing destructive commands. Lucius AI uses two production regions on Cloud Run—Europe and an Australian deployment with customer-managed encryption keys for defense-adjacent customers.
Lucius AI ingests public procurement notices from 13 sources across the UK, EU, US, Canada, Australia, New Zealand, India, Singapore, and World Bank-funded regions. The platform analyzes tender documents using Gemini to generate compliance matrices and draft responses, replacing days of manual review for small and mid-sized suppliers. Retrieval reranking executes directly in AlloyDB with a mean latency of 77 milliseconds, eliminating the need for a standalone reranking microservice.