SABER — System for AI-Based Engagement & Recommendations

Hims & Hers Health

Hims & Hers Health

~2024–2025 Agentic Decisioning
AI Decisioning Warehouse-Native Lifecycle Marketing Hightouch BigQuery ML

Product home (in progress): saber.danieljrobles.com

What SABER Is

SABER stands for System for AI-Based Engagement & Recommendations. It is a self-created, warehouse-native agentic decisioning engine built during my time at Hims & Hers Health—designed to automate lifecycle marketing experimentation and optimization with projected annualized revenue impact.

The Hims prototype proved the architecture with vendor building blocks. A later, independently evolved concept of the same capability lives at saber.danieljrobles.com.

People circuit

Open the live loop

The evolved product watches people move through Rest personas, site events, orchestration waits, activation sends, and engagement outcomes. Color is the system treatment, not the channel. This is a concept demo on this page’s palette; the operator console is linked above.

  • observe
  • wait
  • send
  • measure
  • holdout
  • cancel send
Rest is overall persona (left worse, right better). Coral is a plan cancel after an invalidating action. Green and red KPI ticks live in the operator console.

The Problem

Lifecycle marketing teams need to decide who should receive a message, what that message should contain, and when it should be sent. In a traditional stack, those decisions are spread across warehouse queries, audience tools, campaign builders, and manual review. That makes experimentation slow, ownership unclear, and it difficult to connect an outcome back to the decision that produced it.

SABER treats decisioning as a repeatable system: governed data goes in, a decision is made, the decision is activated through an execution channel, and the resulting signal returns to the data layer for evaluation.

Architecture & Boundaries

The architecture separates intelligence from execution. The warehouse and model layer own features, eligibility, scoring, and decision history. An activation layer translates those decisions into channel-ready audiences and events. The delivery platform sends the message, while response data flows back for measurement and learning.

  1. Data foundation — modeled profiles, events, consent, and eligibility rules.
  2. Audience definition — identify the people and lifecycle moments eligible for a decision.
  3. Decisioning — select the next-best goal, treatment, content, or timing.
  4. Activation — deliver a stable, auditable payload to one or more channels.
  5. Measurement — compare treatment, business-as-usual, and holdout outcomes.
  6. Learning loop — feed results back into models, rules, and future experiments.

Key Decisions & Tradeoffs

  • Warehouse-native intelligence: keep definitions and decision history close to governed data rather than burying business logic in a delivery console.
  • Modular execution: make the execution layer replaceable so the decisioning model does not need to be rebuilt when a channel or vendor changes.
  • Experiment before expansion: begin with a bounded lifecycle use case and define the comparison method before scaling to more audiences.
  • Batch versus real time: use the simplest latency that supports the use case; real-time infrastructure is justified only when faster decisions create meaningful value.
  • Human governance: require reviewable rules, audit trails, QA gates, and clear rollback paths for model-driven decisions.

How It Works

The preliminary build used Hightouch AI Decisioning and Google BigQuery ML infrastructure. SABER runs as a six-stage loop—Hightouch did not cover every stage end-to-end, but the prototype delivered enough to proceed.

  1. Audience segmentation — define who enters the decisioning loop
  2. Goal inference — infer the next-best outcome for each audience
  3. AI-driven content assignment — select creative and offers via models
  4. Multi-channel deployment — activate across lifecycle channels
  5. Engagement feedback loops — capture response signals back into the warehouse
  6. Continuous learning & optimization — retrain and retune from outcomes
Segmentation Goal Inference Content Assignment Multi-channel Feedback Loops Continuous Learning

Technical Stack

Deployed with Docker, GitHub Actions, and CircleCI. An internal UI/UX for operators was self-developed in Cursor so marketing and MarTech users could interact with decisioning workflows without living in raw vendor consoles.

Hightouch AI Decisioning BigQuery ML Docker GitHub Actions CircleCI Cursor Internal Operator UI

Impact & Evolution

  • Defined a six-stage decisioning architecture for lifecycle marketing.
  • Prototyped the architecture with warehouse and activation-platform building blocks.
  • Separated decision logic, activation, and measurement into inspectable stages.
  • Created an operator-facing interface for reviewing and running decisioning workflows.
  • Continued the concept independently at saber.danieljrobles.com.

Results are described at the architecture and prototype level. Employer-specific projections, customer data, and internal implementation details are intentionally omitted.

Lessons for Other Teams

  • AI decisioning is an operating system for decisions, not a single model.
  • Data quality, consent, and experiment design matter as much as model quality.
  • Separating intelligence from execution preserves optionality and reduces migration risk.
  • Every automated decision needs an owner, an explanation, a measurement plan, and a rollback path.