Alexander Jabbour’s work on Project Nebula shows how AI can scale high-quality sales coaching without losing the human voice behind it. By using a dual-agent architecture that learns each manager’s implicit coaching style and applies it contextually, Nebula turns fragmented feedback into personalized, continuous development at enterprise scale. The result reframes AI coaching not as generic automation, but as an authentic extension of experienced leadership.
Alexander Jabbour, currently serving as Engineering Lead at Rilla, is at the intersection of applied AI and sales enablement. With a track record in software innovation and AI-powered product launches, Jabbour specializes in transforming complex concepts into operational enterprise solutions.
His stewardship of project Nebula (also known as “Rilla CoPilot”) – Rilla’s latest sales coaching platform – reflects expertise gained at early-stage ventures and a robust academic foundation in Software Engineering from the University of Waterloo. Industry-wide, limitations in sales coaching coverage have revealed a need for scalable, personalized feedback systems.
As organizations grapple with fragmented evaluation processes and underutilized learning opportunities, new platforms such as Rilla CoPilot are reframing how performance is fostered on distributed sales teams. This trend is propelled by wider adoption of sophisticated language models and hybrid AI architectures, enabling automated, context-sensitive coaching at unprecedented scale and speed.
Project Nebula arose from a recognition that sales managers were overwhelmed by the volume of conversations, making high-quality feedback a logistical challenge. Jabbour observed, “Leaving high-quality feedback on every rep’s call would consume their entire day, time they need for higher-leverage work such as strategy, hiring, and deal support.” He referenced prior success in AI code review, noting the opportunity to adapt similar agent-driven feedback methods for sales environments.
This strategic pivot was informed by direct pain points relayed from users. Jabbour emphasized, “We could solve ‘how do I coach my massive sales team while maintaining quality at scale’ by adapting the same approach my colleague built for code review, but for sales coaching at Rilla.” The move reflects industry findings that coaching interventions, when elevated in quality and frequency, can significantly increase sales goal attainment, as outlined in a mutual insurance study.
Project Nebula’s architecture is structured around two specialized agents: the Research Agent and the comment delivery agent (also referred to as ‘Rilla Copilot’), each fulfilling distinct but complementary functions. “The Research Agent runs a multi-step deep research process using Claude’s Agent SDK with Anthropic’s latest thinking models. It indexes all comments that a sales manager has ever left across their reps’ conversations,” Jabbour explained. This agent surfaces both explicit and implicit coaching methodologies by iteratively reading and documenting behavioral patterns.
The operational layer, the comment delivery agent, then leverages these findings to generate feedback tailored to the individual manager’s coaching voice. “Given a sales conversation transcript and the relevant manager’s knowledge base, it generates feedback that mirrors how that specific manager would coach,” he said.
The architecture includes modes for human-in-the-loop validation as well as fully autonomous coaching, streamlining both quality and scale. This approach aligns with contemporary AI agent trends, such as the integration of powerful LLMs with domain-focused automation, enabled by advances like Claude Opus 4.1.
Building a system that captures and replicates nuanced coaching styles presented several technical obstacles, primarily around authentic synthesis and contextual retention. “There’s something lost in the embedding and retrieval process: the nuance of how a manager phrases things, the implicit rules they follow, the contextual judgment about when to apply which coaching pattern. RAG fragments that context,” Jabbour detailed, addressing the limitations of retrieval-augmented methods.
Instead, Nebula’s agent reads large amounts of raw commentary directly—an approach supported by findings on the capabilities of advanced LLMs for context-rich synthesis. Balancing high-quality, frequent knowledge base updates with the operational costs of large-model runs also posed a challenge, prompting the team to refresh manager-specific datasets on a weekly cadence. This method supports scalable improvement without sacrificing accuracy.
The platform’s technical foundation relies on Claude’s Agent SDK and the latest Claude Opus models, harnessing multi-step reasoning for deeper, systematic analysis. “The heavier model was essential for the depth of analysis required to extract implicit coaching patterns from raw comments,” Jabbour noted.
Foundational design choices, such as storing knowledge bases as structured markdown files, increased legibility for LLMs and streamlined knowledge transfer between agents. He elaborated, “Markdown is highly legible to LLMs, and the structured format we developed – covering practices, rules, conventions, and patterns – was optimized for the comment agent to consume efficiently.”
The deliberate feedback loop in copilot mode allows iterative refinement based on direct user input. Such approaches align with industry advances in multi-file code refactoring and domain-specific model alignment, demonstrating that hybrid architectures and context adaptation are central to enterprise-grade AI deployments.
A hallmark of Project Nebula’s design is its focus on personalizing feedback so that recommendations “sound like” the manager. “Every manager gets their own dedicated knowledge base, there’s no shared model or generic coaching template,” explained Jabbour. In copilot mode, managers review AI-generated comments and provide real-time feedback, directly influencing future outputs.
“When managers review AI-generated comments and indicate whether they’re on target, we use that signal to update the knowledge base,” he added. This closed-loop system ensures the individual coaching voice is continuously refined, avoiding the drift toward generic automation. Current research underscores the necessity of grounding AI coaching in both qualitative and quantitative measures, such as engagement scores and conversion rates, while retaining traceable linkages between activity and outcome.
Guiding an interdisciplinary team through Project Nebula required balancing experimentation and cautious decision-making. “The biggest lesson was simple: slow down and actually listen to stakeholders,” Jabbour reflected. He described breaking complex workflows into manageable, independent components, which allowed team members autonomy without sacrificing context sharing.
“We broke the project into the smallest possible independent chunks. Each component could be owned by an individual. That shared context was easily accessible and could move quickly without blocking others,” he said. The rapid prototyping and validation approach supports scalability and adaptability—qualities recommended in recent literature about LLM agent evaluation and drift management.
Customer feedback has underscored Project Nebula’s effect on time management and perceived authenticity. Jabbour recalled, “One sales director said reviewing calls used to consume her entire Tuesday—now she spends maybe an hour reviewing the AI-generated comments and flagging the few that need adjustment.” The platform thus incentivizes broader coaching coverage while restoring time for strategic tasks.
Jabbour further emphasized, “What validated our approach was hearing managers say the comments ‘sound like me.’ That was the whole bet – that preserving the authentic voice would matter more than generic best practices.” The system also offers reflective insights by surfacing unconscious coaching patterns, echoing research showing that quality coaching practices drive measurable performance improvements across teams.
Project Nebula’s model for expertise capture and replication holds promise beyond sales coaching. Jabbour observed, “Any domain where experienced practitioners give feedback based on pattern recognition they’ve developed over the years is ripe for this approach.” He envisions autonomy settings that enable organizations to determine the balance between suggested and automated coaching, adapting usage to trust and use case maturity.
“The bigger vision is personalization at scale, becoming the default, not the exception,” Jabbour said. This outlook is congruent with predictions for enterprise AI, where scalable, user-specific development programs broaden access to quality training. Technologies such as AI agent integrations and continuous, feedback-driven model loops are increasingly central to digital transformation strategies.
Project Nebula’s architecture demonstrates that AI can augment—rather than replace—human expertise, setting tangible benchmarks for how organizations leverage intelligence to drive performance gains at scale.
At Global Business & Economics Journal, we contribute to the future of business and technology with news, innovative research, expert insights, and thought-provoking analysis.
