Quarterly
Goals Menu
AI engineering priorities for the Queen One platform. Three strategic directions for Maricor to evaluate and prioritize.
Building on the Tile Selector
This quarter, I completed the Tile Selector — an AI model that replaces random tile selection within dragontile flows with intelligent, per-user content matching. Rather than showing merchants' email recipients arbitrary content variations, the model learns which tiles resonate with which customer segments.
Below are three potential directions for Q2, each building on this foundation. They're designed as a menu for prioritization — we can pursue one deeply, two in parallel, or sequence all three across the quarter.
Three Strategic Directions
Each builds on the Tile Selector's intelligence in a different dimension of value.
Tile Gap Analysis & Auto-Generation
Use the Tile Selector's learning to identify where merchants' dragontile flows are underserving their customers. The model already knows which tile types perform well for which segments — we can invert that signal to reveal what's missing.
The first iteration surfaces a gap report: "Your Browse flow lacks product-education tiles for first-time visitors." The second iteration auto-generates tile suggestions to fill those gaps, so merchants only need to click "Approve" to close the loop.
Synthetic Data Generation
As an upstart competing against Klaviyo, Queen One has significantly less historical data to train on. This gap compounds — less data means weaker models, which means slower merchant adoption, which means even less data.
We can break this cycle by using a generalized synthetic data generation approach to create high-quality training data precisely where we're weakest. Targeted synthetic augmentation for underrepresented verticals, customer segments, and interaction patterns could let our models punch above their weight class.
Offer & Discount Integration
Currently, the Tile Selector treats all tiles as if they differ only on content. In reality, tiles vary dramatically on economic proposition — a "10% off" tile versus a "new arrivals" tile aren't just different content; they're different incentive structures with different margin implications.
Incorporating offer mechanics (discount depth, free shipping, BOGO) into our models would let the AI balance engagement against merchant margin. This is a major blindspot today and a significant differentiator if we address it — Klaviyo treats all sends the same way.
Risk vs. Reward
Where each goal sits on the risk-reward spectrum to guide prioritization.
Natural extension of the Tile Selector — makes the existing product smarter and stickier. Merchants get actionable intelligence and eventually auto-generated content. But it's incremental: it makes Queen One better at what it already does rather than opening a new dimension of value.
We're inverting signals the model already produces. Gap analysis is straightforward engineering. Auto-generation can be staged — ship the report first, layer in generation later only if quality is there.
If it works, this neutralizes Klaviyo's biggest structural advantage — years of cross-merchant behavioral data. That's not a feature win, it's a competitive position win. Every model we build going forward gets better. It compounds. For an upstart, solving the data disadvantage is existential-level important.
Synthetic data generation for realistic email engagement patterns is a research-grade problem. Validation is inherently hard — you need the real data you're compensating for to verify the synthetic data is any good. If the distribution drifts from reality, models confidently make worse decisions. Could consume a full quarter with ambiguous results.
Opens an entirely new value proposition that Klaviyo doesn't offer — optimizing for merchant profit, not just engagement. "We make you money, not just clicks" is a sales narrative that wins deals at a different altitude. It also deepens the Tile Selector moat by making it the only system that understands economic trade-offs between content and incentives.
Tractable modeling problem — represent offer mechanics as features and learn their interaction with user segments and margin constraints. Main dependency: whether merchants expose enough margin or COGS data for us to compute profitability. Also requires careful multi-objective calibration so margin optimization doesn't tank engagement.
Let's Pick Our Path
Each direction is scoped to be achievable within Q2. The question is which creates the most leverage for Queen One right now.