Original recording in Spanish, including narration and on-screen text.
An edited system walkthrough using sample data, a synthetic voice, and captions. Robots support the visualization; states and events let you track the work.
33
specialized roles
11
connected departments
Local AI
processing on the device
Owned project · Historical evidence
A technical case: assignments, events and review.
Owned Optimatiza project. This case examines task assignment, recorded states and review; it does not describe a customer production operation.
Initial problem
An assignment needs an owner, a visible state and a reviewable result. A robot animation alone does not establish that the requested work was completed.
Our responsibility
The project brings together a browser interface, task assignment, execution, persisted events and linked outputs for review. The technical case documents how those parts interact.
Design decision
Specialist roles share a local model in the documented setup. NOVA can split an assignment into up to three independent tasks; this is not recursive delegation or a chain that consumes earlier child results.
Architecture and execution decisions
Implementation examined
React and Three.js provide the interface and visual scene; Express handles the server; PostgreSQL stores the records; Ollama supplies the local model. Events are transmitted to the browser through Server-Sent Events (SSE).
An explicit execution boundary
The examined configuration uses one shared qwen3:1.7b model and one concurrent inference. The 33 roles across 11 departments identify specialties, not separate models or simultaneous workers. State history is not model learning.
What was observed, and how to read the evidence
Historical controlled test. NOVA distributed two tasks to Marketing and Finance. Both passed through review and reached the completed state. The audit recorded 39 persisted events and 39 SSE events in matching order.
This supports the task-routing and event-recording mechanism in that test. It does not validate the generated content, customer data handling, business savings or current production operation. No new inference was run for this case update.
The video above combines an initial trial that did not complete the flow with captures from a later successful test, identified on screen. Its edited timing cannot be used to infer latency. The public visual demo helps explore the scene; it does not establish that the local executor is running.
Publishing, sending email, issuing invoices and actions in external platforms require separate integrations and authorization. Robots reflect recorded task states; movement does not prove those external effects.
Structures data received as text and identifies fields needing clarification.
Collections assistant
Proposes follow-ups and messages using the supplied balances.
Reconciliation specialist
Compares records and identifies potential discrepancies to verify.
📊Data and researchTurn information into findings+
Researcher
Summarizes the information provided and organizes findings and questions.
Data analyst
Analyzes information and supports specific calculations for task CSV files.
Report generator
Turns data and findings into clearly structured reports.
💻Development and supportPrepare solutions and tests+
Developer
Prepares proposed code and implementation explanations.
Issue classifier
Organizes issues by type, impact, and required data.
Test reviewer
Proposes test cases and reviews the evidence provided.
💬Customer supportFollow through on inquiries+
Inquiry assistant
Proposes responses using available context and documentation.
Follow-up coordinator
Organizes cases, commitments, and next steps.
Help and knowledge editor
Prepares guides, FAQs, and supporting documentation.
📂Administration and talentOrganize internal administration+
Document manager
Proposes indexes, classifications, and summaries of supplied documents.
Talent coordinator
Prepares onboarding plans and supporting materials.
Internal administration assistant
Proposes agendas, organizes agreements, and prepares follow-up.
🛡️Security and complianceSupport control reviews+
Security reviewer
Identifies potential gaps and controls in the supplied information.
Compliance reviewer
Checks the information against the supplied requirements.
Risk and audit analyst
Organizes findings, risks, and measures for assessment.
Current project scope+
Preparation and review. The current version assigns work, records states, and links results for human review. Roles share a local model, and the executor processes one AI task at a time. The 33 roles do not mean 33 simultaneous executions.
AI on the device. In this configuration, assignments are processed locally without being sent to an external AI API. Protecting data also requires permissions, device security, and integration reviews.
Upcoming connections. Publishing on social media, issuing invoices, recording hours, and taking actions on other platforms require integrations that are still pending. Roles prepare proposals; they do not imply web searches, audiovisual generation, code execution, or automatic audits.
A possible evolution. The same approach could be adapted for human teams to organize tasks and tickets, detect blockers, and distribute workload more effectively. That adaptation is not part of the version shown.
Greater clarity at work
The next improvement starts with a well-organized task.
Let's discuss the process you need to organize in your business.
