US long/short equity manager · $10bn–$25bn AUM · 50–100 staff

Choose the research architecture without closing off what comes next.

Independent technical research helped an investment manager separate the decisions it could make now from the complexity it should add only when the evidence supported it.

Engagement

Focused advisory

A technical working session, targeted research and a detailed written architecture response for the firm's technology and data team.

Shared and private research
2 zones
Technical questions researched
5 tracks
Decisions ordered by evidence
One sequence
Future platform options preserved
No lock-in

The firm was planning an AI-ready research estate at institutional scale. A common pool of external research needed to be searchable across the investment team. Internal notes, models and thesis work needed a separate private layer, isolated at analyst level.

The challenge went beyond choosing a database. The design had to account for extraction quality, retrieval across thousands of companies, source lineage, concurrent use, storage economics and the agents that might later sit above the platform.

The team wanted an independent view before making commitments that would become costly to reverse as the research estate grew.

The architecture sequence

Build the controllable foundation before adding expensive complexity.

The recommendation connected platform design to the research questions, permissions and future agent workloads the system would need to support.

Research architecture decision map

Preserve flexibility while protecting proprietary research

Independent viewPermission awareEvidence led

Two knowledge zones

Shared research

Sell-side reports · common corpus

Private analyst work

Notes · models · thesis material

Non-negotiable control

Shared access cannot expose one analyst's proprietary work to another.

Foundation before complexity

01

Portable raw layer

Keep original documents and source lineage.

02

Metadata + hybrid retrieval

Combine semantic, keyword and structured search.

03

Permissions at query time

Preserve analyst-level segregation in every answer.

Source linked
Cost monitored
Agent ready

Sequenced decisions

Choose now

Ingestion, metadata, lineage and access controls

Validate

Retrieval quality, cost, concurrency and traceability

Defer

Knowledge graph and heavier ML until evidence supports them

Resulting direction

A controlled first phase that tests the architecture in the firm's own environment before long-term commitments are made.

Illustrative architecture sequence. Platform choices remain open until retrieval, permissions, scale and cost have been tested against the firm's real research workflows.

What Altitude assessed

Answer the tradeoffs together, not one platform at a time.

Data foundation

Keep source documents in a portable raw layer, with structured metadata and retrieval services above it.

Platform choice

Start with the platform that fits governed document retrieval while preserving options for heavier ML later.

Search quality

Combine semantic, keyword and metadata retrieval with re-ranking instead of relying on embeddings alone.

Knowledge graph

Prepare entities and source references, then add graph traversal only when a defined query requires it.

Private research

Maintain strict analyst-level segregation through storage, retrieval and every downstream answer.

Economics

Separate storage, indexing and serving so cost, capacity and latency can be observed independently.

A decision framework

Some choices were ready. Others needed to earn their place.

Altitude recommended reliable ingestion, structured metadata, source-linked hybrid retrieval and access controls as the first foundation. Those capabilities were useful under any later platform choice and made retrieval failures easier to diagnose.

A knowledge graph could help with supply-chain, causal and cross-company questions. Adding it immediately would also create an ontology and maintenance burden before the team had shown that ordinary retrieval was insufficient. The recommendation was to preserve the ingredients for a graph, then introduce it for a defined query class.

The same principle applied to the platform stack: use a bounded first phase to test quality, permissions, traceability, expected volume and concurrent cost in the firm's own environment before making a long-term architecture commitment.

What the engagement established

A practical sequence for moving forward without pretending every decision was settled.

The client received an independent, researched view of the architecture, the tradeoffs behind it and the evidence required for the next decision. This engagement covered technical advice rather than a platform build, so no implementation outcome is claimed.

Choosing an AI research architecture?

Bring the use cases, constraints and platforms already under consideration. We'll research the tradeoffs, challenge the sequence and identify the smallest architecture decision worth making next.