AI / Agentic Systems
MSS: Designing a Bounded AI Sales Engine
MSS — Modular Sales System — is an internal multi-tenant AI sales engine I designed to move leads through intake, qualification, conversation, booking, and warm handoff while keeping business truth, workflow state, and safety controls outside the language model.
The problem
Most AI sales demos start with the model and work backward. MSS started with the business constraints: who the system is allowed to represent, what facts are true, what the lead’s state is, what actions are permitted, and when a human has to approve the next step.
My role
I designed and implemented the system around those boundaries: tenant and brand context, workflow state, bounded language generation, approval, integrations, and database enforcement.
Architecture
Business truth lives in structured context. The engine controls workflow. The model controls wording inside bounded rules.
- Lead source
- Intake / scoring
- Workflow state
- Structured business context
- Bounded language generation
- Human approval / safeguards
- Messaging / booking
- Warm handoff
What I built and the decisions behind it
- Buyer → brand → lead tenancy model
- PostgreSQL and row-level security
- Structured business/brand context, locked facts, and gap detection
- Lead intake and scoring
- Deterministic workflow and state handling
- Bounded LLM language generation
- Human approval and confidence gates before outbound sends
- Twilio inbound/outbound messaging and signature validation
- Opt-out handling
- Booking and reminder plumbing
- Warm-lead handoff
- Database-level protections against unapproved sends
- Approval/correction signals for future learning
Verification and send controls
Core tenancy/security, brand context, lead intake/scoring, conversations, language drafting, human approval, Twilio messaging/webhooks, booking/reminder plumbing, and database send enforcement were built and verified.
The documented build requires human approval before outbound sends. Signature validation, opt-out handling, and database-level protections put controls around the messaging path instead of relying on the model to follow instructions alone.
What remains in progress
The core tenancy, context, lead intake, conversation, approval, messaging, booking, and send-control layers were implemented and verified. The deeper learning layer remains scaffolded; it is not a finished self-learning system.
Tools and working methods
- PostgreSQL
- Row-level security
- Twilio
- LLM orchestration
- APIs
- Human approval
Result
An implemented sales-workflow foundation connecting structured business context, conversations, approval, messaging, and booking. This is implementation evidence; no sales or revenue uplift is claimed.
What I learned
My AI work is not limited to prompting. I think about data boundaries, deterministic workflow, permissions, model scope, human controls, integrations, failure states, and how an agent fits into an actual business process.