AI / Agentic Systems
AI is most useful when it has rules, context, and a real job to do.
I design and build AI-assisted systems around real business workflows — sales conversations, content production, customer service, automation, and technical implementation. The model is one component of the system, not the source of truth.
How I think about agentic systems.
Business truth stays outside the model
Important facts, constraints, eligibility rules, and operating context should live in structured data or deterministic logic — not in a model’s memory or imagination.
The workflow controls the model
Use deterministic states and orchestration for the parts that must be predictable. Let the model handle language and judgment only where flexibility is actually useful.
Human approval is a feature
When confidence, brand risk, money, or customer impact matters, human review is a valid control — not evidence that the system failed to be autonomous enough.
Fail closed when the system is unsure
Bad output should stop, route for review, or fall back safely instead of being shipped just because the model produced something.
Capture corrections
Human edits and downstream outcomes are valuable data. Store them so the system can improve without rewriting business truth.
Flagship agentic system
MSS — Modular Sales System
MSS is a multi-tenant AI sales engine designed to take a lead as far toward a booked or sales-ready outcome as the business’s actual process allows, then hand that lead off at the correct boundary.
- Lead source
- Intake / scoring
- Workflow state
- Structured business context
- Bounded language generation
- Human approval / safeguards
- Messaging / booking
- Warm handoff
What I designed and implemented
- 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
Business truth lives in structured context. The engine controls workflow. The model controls wording inside bounded rules.
Implementation status
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.
Content Engine
A local-first, brand-aware content production system built to turn structured creative reasoning and approved source media into repeatable production workflows.
- Brand context
- Topic / formula
- Structured brief
- Validation
- Asset resolution
- Missing-asset generation
- Render
- QA / audit
- Human review
- Delivery
- Feedback
Core implementation
- File-based Node.js orchestration with budget and wall-time limits
- Remotion rendering and resumable Lambda batch production
- Strict structured Brief.v1 JSON contracts
- Schema validation
- Brand/context loading
- Real-footage-first asset resolution
- Missing-asset generation only when needed
- Resumable stages with service-specific concurrency limits
- Per-brand storage and delivery adapters
- Render and output QA / audits
- Human review
- Performance/correction feedback design
Creative reasoning can be flexible. Production execution should be reproducible, inspectable, and able to stop when required inputs or quality checks fail.
Implementation status
The brief, asset, rendering, audit, and delivery paths are implemented, including documented batch ad production. Feedback and rule-proposal components exist, while the complete integrated production-and-learning loop remains in progress.
AI for business operations
DealFlow — Marketplace Resale & Vassa
DealFlow is an AI-assisted operating system I built for a used-phone resale business. It connects marketplace sourcing, device evidence, pricing, seller conversations, inventory, and resale outcomes. Vassa is the conversational assistant; code controls the financial limits and the actions it is allowed to take.
- Marketplace listings
- Identity / condition evidence
- Resale value / buy ceiling
- Permitted next action
- Bounded negotiation
- Owner handoff
- Inventory / resale records
- Outcomes / corrections
What I designed and implemented
- Marketplace discovery, listing-detail capture, and catalog matching
- Photo assessment and provider-backed device-identity checks
- Versioned evidence, conflicts, and source-freshness checks
- Deterministic resale-route valuation, costs, profit requirements, and buy ceilings
- Stateful seller conversations and bounded offer / counter-offer workflows
- Exact-message approvals and action-time revalidation
- Capability controls for observation, questions, outreach, and negotiation
- Inventory, cost-basis, sale, and realized-outcome records
- Vassa web workspace with saved chats, streaming, deal context, and usage visibility
- Bounded MCP tools, immutable action proposals, and an execution audit trail
A good conversation is only useful if the device, price, margin, and next action make sense. The system keeps those decisions tied to evidence and explicit operating rules.
Implementation status
The documented live workflow reached evidence-backed opening offers and bounded negotiation up to an owner handoff. Meeting setup, purchase, and payment remained outside that autonomous scope. The later Vassa web workspace was built and tested; its production deployment is not yet verified. The expanded end-to-end purchase and resale workflow remains unfinished.
Supporting systems
AI voice / customer service
At AAE, I worked on an AI voice customer-service system and personally QA-tested product discovery, pricing, customer-information capture, order creation, and specialist-transfer flows.
AI-assisted development
I use AI heavily as part of my development workflow. My contribution is directing, inspecting, debugging, testing, and shipping technical work with AI assistance, while owning the implementation decisions and QA.
The point isn’t autonomy for its own sake.
The useful question is whether the system can do a real job safely, measurably, and with less manual effort. That’s the standard I use.