Vellum Intelligence:
An intelligence production system.
A multi-agent regulatory and geopolitical intelligence platform with two coordinated cores, specialist agents, a layered deterministic architecture, quality assurance, and human approval gates. The reference build for institutional-grade intelligence production.
Intelligence production at scale, without the failure modes.
Regulatory and geopolitical intelligence is high-volume, time-sensitive, and unforgiving of errors. A single misattributed citation or missed regulatory development can compromise a client relationship.
The conventional approach — analysts manually monitoring sources, writing briefs, reviewing each other — is expensive and inconsistent. The AI-first approach most teams reach for — a single LLM pass over scraped sources — produces output that looks credible but cannot be trusted in production.
Vellum Intelligence was built to resolve this. The architecture treats intelligence production as a layered, schema-driven workflow rather than a single LLM call.
Two cores, specialist agents, deterministic orchestration.
The system runs two coordinated cores:
- The collection core monitors a curated registry of regulators, official journals, court databases, news outlets, and specialist sources. Source-specific collectors handle authentication, rate limiting, parsing, and checkpointing.
- The production core processes collected items through structured stages — relevance filtering, classification, expansion, drafting, and editorial review — each performed by a specialist agent with a clear contract.
Between cores, items pass through a schema-validated handoff. The architecture is deterministic where it matters — orchestration, state, and routing — and probabilistic only where genuine judgement is required.
A defining principle: separation of concerns.
No single agent does too much. Collection agents collect. Classifiers classify. Drafters draft. Editors edit. The orchestration layer routes work between them according to a schema-defined workflow. This is what makes the system auditable, recoverable, and maintainable in production.
Editorial QA and human approval gates.
Every brief passes through editorial QA before delivery. An AI critic agent reviews drafts against quality standards. Human reviewers sign off on briefs before publication. The audit trail captures every step — source, classification, draft, edit, approval — and is retained for compliance.
Where briefs are sensitive or politically charged, additional review gates can be inserted. The architecture supports this without modification because every workflow stage is a discrete, replaceable contract.
Built for institutional-grade environments.
Vellum Intelligence runs on dedicated infrastructure. The deployment model supports on-premise installation, model choice (proprietary or open-weight), and full source-code escrow. For institutional clients with strict data residency or regulatory constraints, this is the architectural starting point — not a retrofit.