I build complete products where AI, workflows, APIs, data and deterministic controls have to work together. My flagship build is PropCommand, a full-stack cross-runtime SaaS in private validation. My applied AI work also spans a deployed bilingual web sales agent, auditable pipelines and multi-tenant data infrastructure.
Case studies
The state label is part of every case: deployed, verified, test-verified, prototype or demo.
Built a complete product surface spanning 34 web pages, 56 API routes and 25 C# files, with multi-tenant data boundaries, deterministic pre-action gates, replay-safe evidence, and release-gated billing and licensing. The latest full receipt records 1,751 passing tests.
The implementation is substantial; the release boundary is equally explicit. Windows/provider conformance, hosted receipts and commercial distribution remain pending.
STATE → PRIVATE VALIDATION · NOT COMMERCIALLY LIVE
Published a public TypeScript/React reference implementation for guarded, idempotent agent actions with structured confirmation and inspectable traces. All scenarios, CRM data and confirmation tokens are synthetic; no production, scale or business-outcome claim.
Interactive demo ↗ · Source code ↗
STATE → PUBLISHED REFERENCE IMPLEMENTATION · SYNTHETIC DATA
Built and deployed a bilingual web sales agent integrated with a dealership CRM; verified the booking path end-to-end with a real contact, appointment and opportunity. Critical booking actions use deterministic checks; an LLM verifier runs on risk-sensitive replies and can veto or rewrite them.
STATE → DEPLOYED · REAL CRM BOOKING PATH VERIFIED END TO END
Processed 4,088 real sales conversations into an auditable diagnostic funnel. The public result uses aggregate counts and does not present the historical diagnostic as a rescue campaign or a sales outcome.
STATE → VERIFIED · 4,088 CONVERSATIONS PROCESSED
Built a Node.js/LLM pipeline that processed 11,837 businesses and narrowed them to 2,135 qualified prospects with personalized briefs.
STATE → VERIFIED · 11,837 PROCESSED → 2,135 QUALIFIED
Built a multi-tenant conversation store with tenant isolation, row-level security and encrypted PII.
STATE → VERIFIED IMPLEMENTATION · NO UNSUPPORTED SCALE CLAIMS
Implemented and test-verified the closed-sale-to-Meta-CAPI event path. The public evidence is limited to the implemented and tested event flow; no downstream performance result is asserted.
STATE → IMPLEMENTED + TEST-VERIFIED
Built and hardware-validated a prototype on a real Android device. The validated scope ends at that real-device prototype.
STATE → HARDWARE-VALIDATED PROTOTYPE
Built a multi-agent control-plane demo with a 57-agent template roster and a four-agent demo floor. Only the four-agent floor is presented as demonstrated behavior.
STATE → DEMO · 57-AGENT TEMPLATE ROSTER · FOUR-AGENT DEMO FLOOR
How I work
A bounded workflow, explicit acceptance criteria and evidence for what actually works.
Map the inputs, integrations, risk boundaries and acceptance criteria before expanding scope.
Integrate the agent or automation with the real stack, then add deterministic checks, evals and useful observability.
Record what is deployed, verified, test-verified or still a prototype, then hand off the system and its operating limits.
Work with me
Open to contract and full-time remote AI automation work. Spanish native / English fluent. US Eastern time.
Stack
Tools used across the evidence-backed work above and adjacent prototypes.
Public code: agent-integrity-console · video-grabber · seevideo