Atlas — internal intelligence for a regional investment office

Brief

A regional family office managing diversified assets across the GCC had accumulated thousands of internal memos, deal documents, market analyses, and regulatory filings. The institutional knowledge was real. It just wasn't reachable. Junior analysts spent hours retrieving context that senior partners carried in their heads. Critical decisions waited on briefings that took days to compile.

The brief was to build a private intelligence layer: an AI system that could read everything the firm owned, surface what mattered, and answer questions in natural language with citations back to source. Not a chatbot. An analyst.

Approach

We started with an audit of the firm's information estate. SharePoint, internal Google Drive, email archives, third-party data subscriptions, regulatory feeds. Every source had different document types, different access rules, different update cadences. The architecture had to handle ingestion, parsing, retrieval, and access control at scale, while feeling immediate to use.

We designed the system in four layers:

Ingestion. Connectors to each source. Document parsing with OCR for scanned files and layout-aware extraction for tables. De-duplication. Preservation of source-level access rules so the system never grants visibility the source system didn't.

Knowledge. Chunking with semantic overlap, embeddings generated via OpenAI's text-embedding-3-large, vector storage in pgvector on managed Postgres. Hybrid retrieval combining BM25 keyword scoring with semantic similarity — higher precision than either alone.

Reasoning. Retrieval-augmented generation layer with full citation chain. Claude Sonnet for fact-grounded synthesis. Query routing to specialised agents for compliance, valuation, or due-diligence questions.

Interface. Desktop conversational app for analysts. Mobile companion for partners. Auto-generated daily briefings pushed at 06:30 GST.

The hardest problem was access control. Different documents had different visibility. Deal-room access for partners only. M&A drafts gated by NDA. Regulatory data shared with compliance. We built row-level security on Postgres and per-document ACLs that the retrieval layer respects before any model sees the content. No partner sees what they shouldn't. The model never receives content the requesting user can't access.

What was built

• Ingestion pipeline across six source systems, processing 47,000 documents on initial load
• pgvector knowledge base with 2.1 million embeddings, sub-200ms hybrid retrieval at p95
• Reasoning layer with Claude Sonnet, full inline citation chain for every response
• Desktop conversational app (React + Electron) with saved queries, thread history, report export
• Mobile briefing app (React Native) — partners receive a 90-second AI brief every morning
• Weekly auto-generated digests covering portfolio companies, market signals, and regulatory changes

Outcomes

• Average time from query to cited answer: 8 seconds
• Analyst preparation time for investment memos: reduced 64%
• Daily active partners: 11 of 12
• Regulatory updates surfaced automatically per week: ~40
• Documents queried at least once within first month: 92%

Tech stack

• Postgres with pgvector extension
• OpenAI text-embedding-3-large for embeddings
• Anthropic Claude Sonnet for reasoning
• TypeScript / Node.js backend
• React + Electron desktop, React Native mobile
• Row-level security with per-document ACLs
• AWS, single-tenant deployment

Closing

Atlas was the studio's first AI intelligence build at this scale. The pattern — secure ingestion, hybrid retrieval, model-aware access control, conversational interface — is the foundation we now apply to similar briefs across finance, government, and professional services.

Status: Live since Q3 2025. Three additional verticals in development.