## Overview and Business Context Mowie is an AI marketing platform founded by Chris Okconor (CEO and technical lead) and Jessica Valenzuela (co-founder handling customer experience and commercial operations). The company emerged from the founders' previous eight-year experience running a customer data management startup serving small and medium businesses (SMBs). During that time, they observed a critical pattern: out of 45 customers, only two were actively using the marketing platforms they had paid for, despite these being basic tools for customer segmentation and email marketing. Business owners frequently asked Jessica for marketing advice and assistance that was completely outside their company's scope, including requests to actually create marketing content and manage campaigns for them. The target customers are primarily small to medium-sized businesses in three categories: restaurants, retail, and e-commerce/direct-to-consumer brands. These businesses face unique scaling challenges—a single-person marketing team might manage eight different restaurant concepts with three to ten locations each, effectively making them enterprises despite being classified as small businesses. The time constraints and expertise gaps make sophisticated data-driven marketing essentially impossible for these teams to execute consistently. ## Technical Evolution and Architecture Mowie's technical architecture evolved through several distinct phases, exemplifying an iterative "one bite of the apple at a time" approach to building LLM-powered products. The founders began by having Jessica manually perform what would eventually become the AI's role—creating segmentations, automations, and reading data insights to determine appropriate campaigns. This hands-on experience proved invaluable for understanding the problem space and establishing ground truth for what good marketing execution looks like. The first technical prototype focused on using generative AI to answer questions based on the Simon Sinek "Golden Circle" framework (the why, what, and how of a business). Chris built automated workflows to infer these core business research questions by pulling information from third-party sources and point-of-sale systems, then generated PDFs for customers. This initial version was essentially a series of simple workflows where different LLM calls would find different pieces of information that were then aggregated. Early experiments with image generation for marketing content revealed the limitations of the technology at the time (approximately one to two years ago from the interview). The team encountered now-classic generative AI failures like "eight-finger turkeys" and floating cupcakes in front of the Eiffel Tower. However, through extensive testing and experimentation with different angles, plates, and scenes, they eventually achieved the ability to automatically create different product shot angles from a single photograph, enabling seasonal content without requiring new photography or production. ## Document Hierarchy as LLM Context Architecture The core innovation in Mowie's architecture is what Chris describes as a "hierarchy of documents" that serves as the context layer for LLM-powered content generation. This represents a sophisticated approach to managing the context problem inherent in LLM applications. The founders explicitly rejected highly structured schemas at every level after those approaches "failed spectacularly" in terms of output quality and reliability. Instead, they adopted a loosely structured approach using markdown as the primary intermediate format, with only a few critical artifacts (like calendars and product mixes) requiring precise structured output. The document hierarchy consists of approximately 80 different artifacts that Mowie analyzes to create what they call a "brand and business dossier" and an "audience dossier." These documents are organized hierarchically such that changes to higher-level documents trigger regeneration of dependent lower-level documents. Each document in the hierarchy has its own refresh cycle (typically weekly or monthly by default), and the system employs fingerprinting to detect meaningful changes—if a document regeneration results in less than 5% change, it may not trigger downstream updates. The brand dossier presents customers with several key sections: business profile, customer intelligence, competitive analysis, customer reviews, catalog intelligence, industry trends within local markets, sales and marketing intelligence, and marketing pillars. The marketing pillars specifically define the content strategy, including the mix of educational content, promotional content, and brand awareness content that's appropriate for that business type. This architecture elegantly solves the problem of providing relevant, business-specific context to LLMs while maintaining traceability and transparency. When Mowie recommends a specific campaign or post, customers can trace back through the document hierarchy to see exactly which customer reviews, sales data, or competitive intelligence informed that recommendation. ## Onboarding and Data Ingestion Mowie's onboarding process is deliberately minimalist from the customer's perspective: they provide only their business name and website URL through a form. From there, Mowie ingests publicly available data across the internet, analyzing the approximately 80 artifacts mentioned earlier. The system looks for web presence, product or services catalogs, third-party articles and magazine mentions, and customer reviews from platforms like Google and Yelp. The ingestion process leverages both LLM-friendly APIs and traditional web scraping. The founders note they're fortunate that many review services and point-of-sale systems now offer API access that's relatively easy to work with. Most of the API calls are fixed regardless of the business—for example, retrieving the last 500 reviews from a location. However, some queries use templates that vary based on business type and location, such as searching for relevant sports events that differ based on whether a location is in a football-focused region versus hockey-focused region. A critical insight is that this process doesn't always work perfectly, and Mowie builds error handling and human-in-the-loop checkpoints into the flow. Common issues include business name ambiguity (multiple "Joe Cigars" locations in different regions), multiple location handling for chains, and resolving closed locations that still have active Yelp pages. The system alerts customers to these ambiguities during onboarding and provides interfaces for customers to correct and clean up their own data. ## Content Calendar Generation and Templating The content calendar generation represents the output layer of Mowie's LLM workflow. The system operates on multiple time horizons: quarterly for overall strategy and major campaigns, weekly for detailed content planning, and nightly for determining what specific posts need to go out the next day. The calendar synthesis draws from three distinct calendar streams: public/well-known events (national, regional, or local like the Super Bowl or SF Beer Week), business-specific events that the company has published themselves (like a wine bar's regular Friday night blues open mic), and longer-term quarterly campaigns associated with the business's customer segments. The system weights these based on customer segment data—if a restaurant's customers don't align with beer-focused events, SF Beer Week won't rank highly in recommendations even if it's a popular local event. The actual content generation uses templated prompts with structured inputs. Chris describes it as primarily workflow-based with established prompt templates that incorporate three main pieces of context: general business context from the document hierarchy, the specific marketing pillar being addressed (educational, promotional, brand awareness), and the target channel (Instagram, Facebook, email, etc.). The system generates content by populating these templates: "We're generating an educational post for Instagram" with the business context informing what that educational content should specifically address. The templates include three main override slots where businesses can provide input: the specific product being featured, tone adjustments, and a more open-ended agentic override where customers can provide general input that modifies the overall content creation prompt. This gives customers control while maintaining automation as the default. ## Human-in-the-Loop Design Mowie's human-in-the-loop design is sophisticated and multi-layered, recognizing that different stakeholders need different levels of interaction. The weekly workflow for customers is intentionally lightweight: they receive an email summarizing the upcoming week's content and can approve it directly from the email. Posts are queued ahead of time and can be rolled back if a customer later realizes something has changed in their business. Customers can also log into the platform to view their full content calendar, override specific posts, and regenerate content by modifying the context inputs. During the quarterly review cycle, customers see recommended campaigns ranked by relevance to their business and customer segments. They can approve Mowie's top recommendations or choose different events from the full list. The system tries to limit businesses to three main calendar campaigns and three to five specific events per quarter to avoid overwhelming them. The founders explicitly recognized that asking customers to trace through context hierarchies to debug problems is like "telling them to eat broccoli"—it's good for them but they often won't do it. They've implemented lightweight feedback mechanisms including thumbs up/thumbs down buttons, the ability to highlight and circle problematic UI elements or data, and immediate in-app feedback when customers encounter content that doesn't align with their brand. This balances the need for structured feedback with the reality that bug report workflows are too cumbersome for busy small business owners. Interestingly, they've found that certain personality types—particularly business owners with engineering backgrounds or those who naturally enjoy spreadsheets and data—do engage deeply with the transparency features and provide valuable feedback on the underlying inferences. ## Evaluation Strategy Mowie's evaluation strategy operates at multiple levels, reflecting the complexity of their LLM-powered workflow. At the highest level, they embrace "the customer as the ultimate eval"—if customers accept content without edits, that's a strong positive signal. The system tracks what customers accept without changes, what they edit, and what they regenerate, using these signals to understand what's working. For content performance, Mowie integrates with the underlying sales systems (point-of-sale for restaurants, e-commerce platforms like Shopify) to measure actual business impact. They acknowledge that attribution isn't perfect—there isn't an exact link between a specific post and sales—but they work to establish statistically meaningful inferences about how marketing impacts product sales or sales within customer segments. Posts are monitored for performance, and if content isn't performing well after 48 hours, the system can automatically stop running it as an ad and recommend new content. Traditional engagement metrics (opens, clicks, dwell time) are tracked but the founders express more excitement about the actual purchase journey—did the click-through actually result in a sale? This focus on business outcomes rather than vanity metrics reflects their deep understanding of what small business owners actually care about. For the document hierarchy and business understanding layer, Jessica initially served as human-in-the-loop validation for every step during the pilot phase. Now the system monitors whether businesses view specific documents (indicating areas of interest or concern), whether they make edits to documents (particularly the more inference-heavy ones like customer segmentation), and when they override recommendations at the calendar or content level. When customers override recommendations, they're shown the context that went into that recommendation and asked to indicate what's working well or what needs help through simple thumbs up/thumbs down interfaces. Chris and Jessica still manually review calendars being produced across their customer base to validate that the system is working correctly and to catch issues that customers might not report. This ongoing manual review serves as a meta-evaluation layer and helps them understand whether the document hierarchies are providing sufficient context and whether new information is being incorporated appropriately. ## Technical Challenges and Architectural Decisions One of Chris's key technical insights was that attempting to define strict structure at every level of the document hierarchy failed. The loosely structured markdown approach with selective structured outputs only where necessary proved more reliable for LLM inference quality. This reflects a broader pattern in LLM applications where over-constraining outputs can paradoxically reduce quality. The ambiguity resolution during data ingestion represents another significant challenge. Business names, multiple locations, and inconsistent data across review platforms require sophisticated error handling and customer interaction flows that go beyond simple data retrieval. The founders also grappled with the evolution from tool assistance to full workflow automation. They started by using AI to help Jessica's workflow be more efficient—word-smithing, content creation assistance, image generation for reusing assets. As they automated more pieces, they realized they needed to support an end-to-end workflow rather than providing disconnected tools. Customer feedback consistently pushed them "upstream"—from helping create posts to suggesting what posts to create to building entire content calendars to defining marketing strategy. Chris mentions that from an architectural perspective, they simplified somewhat by not generating calendars in real-time. The quarterly/weekly/nightly cadence structure allows for more sophisticated processing and quality control compared to fully real-time generation. ## Multi-Channel and Attribution (Future Directions) The founders identify attribution as a major focus area for completing their product vision. They want to provide customers with a clear picture of where advertising dollars are being spent, which channels are most effective, and what messaging resonates with different customer segments and personas. They're also exploring omnichannel expansion beyond digital social content and email. Customer use cases are driving this—some customers are already using Mowie-generated social content on their websites for operational content refreshes. Others in the digital out-of-home advertising space (screens in stores, billboards outside venues) are exploring how to use Mowie's automated, segment-targeted content for channels traditionally limited to brand awareness and static messaging. The vision is for a local business to have a digital billboard outside their retail store that updates based on time of day, day of week, and current campaigns—effectively running pumpkin spice latte promotions on their own dynamic signage. ## Business Model and Product-Market Fit Signals While not extensively discussed in the interview, several signals point to product-market fit. The founders grew from manual concierge service with early customers to building the product based on demonstrated demand. Customer feedback consistently requested more automation and more upstream involvement in marketing strategy. The willingness of businesses to trust Mowie enough to approve content via email without extensive review suggests the quality meets their needs. The founders' domain expertise—eight years working hands-on with restaurants, retail, and e-commerce businesses on marketing and customer data management—clearly informed their product decisions and gave them credibility with customers. Jessica explicitly served as the "physical Mowie" before the AI version existed, giving them ground truth for what good marketing execution looks like in this context. ## LLMOps Lessons and Patterns This case study illustrates several important LLMOps patterns. The document hierarchy as context architecture provides a reusable pattern for managing complex business knowledge in LLM applications. The fingerprinting approach for detecting meaningful changes and triggering regeneration offers a practical solution for keeping LLM-generated content fresh without excessive computation. The evolution from structured schemas to loosely structured markdown reflects hard-won lessons about LLM reliability and quality. The multi-layered human-in-the-loop design shows how to balance automation with customer control and trust. The focus on business outcome metrics (sales) rather than just engagement metrics demonstrates mature thinking about evaluation. Perhaps most importantly, the iterative "one bite at a time" approach—starting with golden circle inference, moving to basic social posts, then to calendar generation, then to full strategy—provides a roadmap for other teams building complex LLM-powered products. The founders didn't try to build the complete vision at once; they found the minimum valuable automation, shipped it, got feedback, and expanded based on what customers actually needed.