Product Context
Aimpoint Digital is deploying a predictive lead-prioritization solution into a client’s existing sales and operations workflow built on Salesforce and downstream BI dashboards. End users are sales reps, sales managers, and RevOps teams who need ranked leads and next-best-action recommendations embedded directly into their daily process.
Scale
| Signal | Value |
|---|
| Client DAU (internal users) | 18K sales and operations users |
| Leads/opportunities in CRM | 45M historical, 2.5M active |
| New lead events/day | 1.2M |
| Peak scoring QPS | 3.5K requests/sec during business hours |
| Batch rescoring volume | 2.5M active records every 4 hours |
| p99 latency budget | 250ms end-to-end for synchronous scoring |
| Regions | US + EU deployments |
Task
Design the end-to-end ML system and deployment lifecycle for introducing this AI solution into the client’s operational workflow.
- Clarify the product requirements, user journeys, and integration points across Salesforce, APIs, and Aimpoint Digital reporting surfaces.
- Propose the full architecture from data ingestion and feature computation through candidate generation, ranking, re-ranking, and delivery into the workflow.
- Define the training, validation, deployment, and rollback lifecycle, including how you would avoid training-serving skew and manage feature drift.
- Specify online vs. batch inference paths, latency budgets, and capacity planning at the given scale.
- Describe offline and online evaluation, experimentation, business guardrails, and post-launch monitoring.
- Identify major failure modes, including operational, model, and workflow-adoption risks, and explain mitigations.
Constraints
- Predictions must be visible inside the client’s existing Salesforce workflow; reps should not need to switch tools.
- Some features arrive in near-real time from event streams, while others are refreshed only daily from ERP and marketing systems.
- The system must support explainability for account teams and auditability for regulated client environments.
- EU client data must remain region-local, and PII usage is restricted.
- Cost target: keep average inference cost below $0.002 per scored record.
- The client expects graceful degradation: if the model is unavailable, rules-based prioritization must continue.
