AI advisory · How Urdima works
For technical readers, security reviewers and anyone who wants to see what is actually running. Every component states whether it is operational, in rollout, in development, or planned.
Six layers around the core
Capability registry
A structured registry describes every agent, skill, integration and workflow: what it does, when to use it, what it pairs or conflicts with, and what governance floor it needs, and routes each request to the right specialist. New capabilities pass a structured review before becoming routable.
What is running
Work is not manually assigned to a tool. A structured registry describes every agent, skill and integration: what it does, when to use it, what it pairs or conflicts with, and routes each request to the right specialist. Read-only by design: it recommends, never modifies project state.
For decisions where one model's judgement is not enough, three independent models from three providers review the same brief in parallel and a fourth reconciles them, surfacing disagreement rather than hiding it. Panel output is advisory, always independently checked.
Every project carries durable objectives, a map of workstreams, a live task list and an interrupt log, enforced so state survives crashes, model switches and multi-day gaps. Any fresh session can reconstruct where a project stands.
A local semantic index over more than ten thousand project documents, refreshed twice daily and queried from within any session, so ‘what did we decide about X’ is answered from the record, not from memory.
From a phone, by text or voice, in English or European Portuguese: query any project's state, search the index, resume a session where it left off, direct governed work through a private, non-federated channel never exposed to the public internet.
A daily cross-account brief consolidating and prioritising messages; index refreshes; one-off follow-ups queued for a future date that fire unattended. Verified run history, not configuration on paper.
A private on-premise stack running open-weight models for drafting, reasoning, vision and embeddings, with no data leaving the machine, for any context requiring zero cloud exposure.
The verified-answer pipeline is in rollout: built and partially verified. It is designed to link factual claims to source citations, use a separate verifier to check the primary model's answer, check calculations deterministically and record verification steps for audit.
A per-item multi-lens review during production, and a multi-cycle batch audit for large artifact sets: calibrate, target, sweep, gate. No plan is approved on a single pass; every step names an executor and a distinct reviewer.
A portable executive-productivity interface on the same engine, extended with mobile access, an inbox and ideas-routing layer, and richer project state. In development, served privately, ahead of formal release.
Multi-model review
For architecture decisions, strategic plans and contested claims selected for multi-model review, models respond separately to the same brief in parallel. A synthesising model compares their positions and highlights disagreements for review.
Security, privacy and data residency
Accuracy and verification
Current maturity
Contact
A few lines are enough. We reply within two working days.
2F Advisory · Lisbon, Portugal
