Decision Agent
An AI judgment layer that helps D2C ecommerce founders convert multi-tab merchant data into timely ad spend decisions.
D2C founders routinely miss short, high-ROI growth windows or burn capital advertising out-of-stock products, because performance dashboards only report data without providing the judgment layer needed to confidently commit money in the moment.
A dashboard reports what happened, but it doesn't explain why. Founders end up juggling between disconnected browser tabs (Meta ROAS, Shopify inventory, bank balances, etc.) and complex data boards, causing decision fatigue while fast-closing growth windows slip away.
Ad Compass analyzes ROAS, inventory days-of-cover, and cash-cycle gaps to deliver timely recommendations, providing an interactive slider to modify spend, a one-tap authorization rail, and an inverse guardrail logic that prompts ad pauses in instances like “inventory drop below a 3-day cover window.” To validate the agent's behavior, I engineered a 7-day time-stepped ledger simulation that models real-world D2C ad spend and user behavior. I built the sandbox in React, TypeScript, and Tailwind on Lovable, with deterministic math modules in the frontend calculating rolling CPC and CAC to shield the AI layer from arithmetic errors. To mitigate hallucination and prevent context spillage across the timeline, I designed data-grounded prompts, time-slice context filters, and structured JSON parsing boundaries around Gemini 3.1 Flash Lite.
I could have built autonomous auto-transfers, multi-channel marketplace integrations (Amazon, Flipkart), and live Account Aggregator (AA) bank consent APIs. I deliberately cut all three. Auto-movement destroys founder trust, multi-channel early expansion dilutes attribution signal, and upfront API consent adds friction before value is proven. Constraining the MVP strictly to the Meta-to-Shopify funnel with a 1-tap manual override aims to prove the core judgment loop with zero operational risk.
North Star: Recommendation Acceptance Rate (% of AI-suggested deployments authorized by the founder within 12 hours). Directly measures whether the agent’s financial judgment is trusted enough to move real money.
Supporting: Time-to-Deployment (average hours elapsed between the dashboard detecting a performance signal and the capital being routed). Validates the core hypothesis that we are successfully collapsing the decision-making bottleneck.
Counter: The "Skip/Dismiss" Rate. If founders frequently ignore the alerts or constantly use the slider to scale the recommended amount down to zero, the AI's risk-weighting is failing, creating alert fatigue rather than utility.
I would run a 30-day "shadow testing" phase with a closed cohort of 10 D2C design partners. The agent would ingest their live data and generate recommendations, but we would strictly measure those suggestions against what the founders actually chose to spend in reality. This allows us to tune the risk algorithm safely before touching real capital. Once the core math proves reliable, the vision is to scale Ad Compass into the default decision layer for how D2C founders route ad capital across networks.
This case explored: