An enterprise knowledge agent that remembers

The Challenge

Lumina’s internal research agents were drowning in 35,000+ documents — shipping manifests, compliance policies, carrier contracts, and years of operational reports. Analysts were spending 4–6 hours per day just finding the right context.

Key Pain Points:

• 62% accuracy on complex research queries

• Researchers wasting 4+ hours/day on document search

• Inconsistent answers depending on which agent handled the request

• No versioning — agents often cited outdated policies

• Compliance team constantly auditing “where did this answer come from?”

The agents had access to documents, but no real memory or intelligence layer.

The Solution

We built a production-grade Intelligent Knowledge Base with:

• Semantic + metadata hybrid search across 35k+ documents

• Automatic versioning with policy expiration awareness

• Smart chunking and metadata enrichment (carrier, region, document type, effective date)

• Continuous learning loop that improves retrieval based on analyst feedback

• Full audit trail for every answer (source document + timestamp + version)

The system was designed so any of Lumina’s 12 research agents could instantly pull the correct, up-to-date context.

The Results (After 10 Weeks)

Quote from Lumina’s Director of Operations

Automat gave our agents a brain that actually remembers. Our analysts went from spending half their day digging through PDFs to getting accurate, sourced answers in minutes. The compliance team is finally happy too — every answer now has a perfect audit trail.

What Changed for Lumina

Before Automat:

• Every research agent was essentially a fancy search bar.

• Analysts had to manually verify every answer because they didn’t trust the AI.

• New analysts took 3–4 months to become productive because they had to learn the document landscape.

After Automat:

• Agents surface the exact right document + version + relevant clauses in <3 seconds.

• Analysts trust the output and spend their time on high-value analysis instead of search.

• New hires become productive in weeks because the memory layer does the institutional knowledge transfer.

Technical Highlights

• Architecture: Hybrid vector + knowledge graph with temporal awareness

• Chunking: Semantic + structural (tables, sections, appendices preserved)

• Versioning: Automatic policy expiration detection + “effective date” awareness

• Feedback Loop: Analyst thumbs up/down directly trains the retrieval model

• Latency: <220ms even on complex multi-document queries

What’s Next

Lumina is now expanding the system to their carrier onboarding agents and customs compliance agents. The same memory infrastructure is being reused across 8 different agent teams.