It's like having a tireless executive assistant who juggles everyone's calendars across email, text, and WhatsApp and never double-books anyone.

Vela's core scheduling engine frames every coordination request as a dynamic constraint satisfaction problem (CSP), where participants' availability windows, timezone differences, priority levels, room/resource constraints, and channel preferences are all modeled as variables and constraints. The system ingests real-time calendar data from Google Calendar and Outlook, parses natural language requests via LLM-powered NLP (e.g., "sometime early next week, but not Monday morning"), and continuously re-solves the CSP as new information arrives—cancellations, new participants, or shifting priorities. When conflicts arise, the agent autonomously negotiates alternatives across email, SMS, WhatsApp, and phone, applying priority weighting so that VIP or revenue-critical meetings take precedence. A reinforcement learning layer learns from historical outcomes (e.g., which proposed times get accepted, which channels yield faster responses) to improve proposal quality over time. The graph database (forked KuzuDB) maintains persistent relational context—who has met before, recurring preferences, organizational hierarchies—enabling the agent to make increasingly intelligent scheduling decisions without human intervention.

It's like an air traffic controller for calendars—except instead of planes, it's juggling your CEO's investor call, three panel interviews, and a dentist appointment, and nobody crashes.

It's like having a politely persistent friend who knows exactly when and how to nudge someone who hasn't replied—without being annoying.

One of Vela's most novel ML applications is its adaptive follow-up and ghosting recovery system. When a scheduling thread stalls—a participant stops responding, ignores a proposed time, or goes silent after an initial reply—the AI agent doesn't simply resend the same message. Instead, a reinforcement learning model evaluates the context: how many times has this person been contacted, on which channel, at what time of day, and what is their historical response pattern? The model then selects the optimal next action from a policy space that includes switching channels (e.g., moving from email to SMS), adjusting message tone and urgency, varying send timing, or escalating to a human coordinator. Each outcome (response received, meeting confirmed, continued silence) feeds back into the reward function, continuously refining the agent's follow-up strategy. The system also leverages the KuzuDB graph to understand relational context—if the unresponsive person is a C-suite executive, the agent may route through an executive assistant node in the relationship graph rather than contacting them directly. This creates a self-improving loop where Vela gets measurably better at recovering stalled threads over time, directly impacting pipeline velocity for sales teams and fill rates for recruiting organizations.

It's like a golden retriever that learns whether you respond better to a bark at the front door, a nudge on your hand, or just sitting next to you looking sad—and picks the right one every time.

It's like giving the AI a photographic memory of every meeting, preference, and relationship so it never asks you the same question twice.

Vela's engineering team has forked KuzuDB, an embedded graph database, to build a persistent agent memory layer that tracks the full relational context of every scheduling interaction. Unlike traditional scheduling tools that treat each request as stateless, Vela's graph stores entities (people, teams, organizations, meetings, channels) and their relationships (reports-to, prefers-morning, previously-met, VIP-status, timezone) as nodes and edges. When a new scheduling request arrives, the agent queries this graph to retrieve relevant context: Has this participant been difficult to schedule before? Do they prefer SMS over email? Are they in a different timezone than last quarter? Is there an executive assistant who should be looped in? This context is injected into the LLM prompt and constraint solver, enabling decisions that reflect accumulated organizational knowledge rather than just raw calendar availability. The multi-writer concurrent architecture allows multiple scheduling agents to update the graph simultaneously without conflicts—critical for enterprise deployments where dozens of recruiters or sales reps may be scheduling in parallel. Over time, the graph becomes a rich, queryable knowledge base of an organization's scheduling DNA, powering features like automatic preference learning, relationship-aware routing, and predictive availability modeling.

It's like the difference between a new temp who asks "Who's your boss?" every Monday and a seasoned office manager who already knows everyone's coffee order, parking spot, and which two people should never be in the same room.