Lucius AI has moved its tender search platform to Google's AlloyDB for PostgreSQL, cutting semantic search query times from 1.14 seconds to 24 milliseconds, according to the company.
The startup, which tracks public procurement markets across five continents, now runs its relational data, audit logs and vector embeddings in a single database system. The setup supports a catalogue of more than 210,000 tenders from procurement sources in the UK, Europe, North America, Asia-Pacific, Africa and Asia.
Its platform is aimed at businesses bidding for public contracts. Lucius AI uses Google's Gemini models to analyse tender documents and produce compliance matrices, bid recommendations and draft responses that cite original source pages.
For the solo founder, the database consolidation is also an operational decision. Lucius AI manages nightly ingestion from 13 public procurement sources, along with analytics, performance tuning, data validation and incident response.
Database shift
Instead of splitting workloads across separate relational, vector and logging products, Lucius AI has consolidated them in AlloyDB. This keeps vector embeddings alongside relational records while using a single backup schedule and identity management framework.
Authentication runs through Cloud IAM, with service accounts mapped to database roles. Lucius AI does not store database passwords in application environments, and AlloyDB provides automated backups and point-in-time recovery.
The platform runs in two production regions: one deployment in Europe and another in Australia. It also uses a separate AlloyDB cluster and customer-managed encryption keys for defence-adjacent customers.
Search speed
The drop in query latency followed Lucius AI's move from unindexed vector comparisons to a ScaNN index in AlloyDB. The change was based on an automated performance audit by an AI agent connected to the database.
The same database environment is used for retrieval reranking through the ai.rank function. Mean latency for reranking was 77 milliseconds, removing the need for a separate reranking service.
Lucius AI also said rebuilding its semantic index involved embedding 115,820 records in 10.6 minutes using the Gemini embedding model, at an API cost of about USD $3. AlloyDB auto embeddings now keep those vectors current.
AI oversight
Lucius AI has connected an AI agent to the database through the Model Context Protocol, or MCP, to automate routine administrative work. It uses the open-source MCP Toolbox for Databases with the prebuilt alloydb-postgres server.
Access is tightly limited. The agent uses a dedicated PostgreSQL role with SELECT access across the schema and UPDATE permission on one operational table, while destructive commands such as DROP, DELETE and TRUNCATE are excluded.
Under that model, the agent handles on-demand analytics, query-plan inspection, index analysis, incident forensics and daily data-quality checks across all 13 procurement sources. In one case, Lucius AI said the agent analysed audit logs after an external security probe and reconstructed the request timeline within minutes.
Lucius AI said this approach reduces the need to build manual dashboards or maintain separate analytical pipelines. It also argued that a progressive permissions structure lets companies start with read-only access before granting wider rights where needed.
Operational load
Lucius AI's business model depends on processing large volumes of public procurement material with limited staffing. Its coverage includes the UK, the EU, the US, Canada, Australia, New Zealand, India and Singapore, as well as World Bank donor-funded notices across Africa and Asia.
That breadth creates a substantial operational burden for a small organisation, particularly when customers expect current tender data and secure handling of commercially sensitive searches. Lucius AI said managed database services combined with AI-led database operations are intended to keep that burden low without adding dedicated data engineering or database administration teams.
Lucius AI is also using AlloyDB's columnar engine with auto-columnarisation. According to the company, the database identified and stored 40 frequently queried columns across four tables in memory within a day, improving reporting queries without requiring a separate analytical store.
It has validated AlloyDB AI embedding functions across the full catalogue and now uses a weekly maintenance job to refresh vectors. Lucius AI said the MCP-based operating model works best when destructive actions remain restricted to human administrators.