Diego González Zapiain

Work /

PAZZ Marketplace

A vehicle-leasing marketplace in Mexico. As Technical Lead, I'm rebuilding its platform as an agentic AI stack, so discovery, quoting, qualification and partner handoffs work as one path.

Technical Lead · May 2026–present · Mexico · Private commercial platform

In development

PAZZ quote composer with an Audi A5 Coupé TFSI Quattro spec and a four-step progress bar
Specify the vehicle. A vehicle spec for an Audi A5 Coupé TFSI Quattro (2025), listed at $1,200,000 MXN and quoted against $500,000, with a four-step progress bar and a running vehicle summary.
PAZZ fiscal-regime question with four branches and a lender/SOFOM disclaimer
The fiscal question. A four-way fiscal-profile branch — empresa, por mi cuenta, asalariado, no seguro — with leasing explained against credit, and the disclosure that Pazz is not a lender or a SOFOM.
PAZZ real-cost estimate showing rent, estimated ISR deduction, and creditable IVA at 48 months
Real cost, not just rent. Rent $14,263, an estimated ISR deduction of −$1,800 and creditable IVA of +$960, for a real cost of $12,463 at 48 months — an illustrative estimate, not a lessor offer.
PAZZ quote result comparing 48, 36, and 24-month terms with initial payment and purchase option
The quote. A 48/36/24-month term comparison, with an initial payment of $44,030 and a $217,500 purchase option.
PAZZ Claudio entry screen disclosing it as a virtual adviser
Ask Claudio. Claudio's entry screen, disclosed in-product as a virtual adviser that does not decide applications.
PAZZ partner application intake with synthetic details and a revealed Razón social field
Partner intake. A partner application filled with synthetic applicant details; checking Persona Moral reveals the conditional Razón social field. This surface is flag-gated and off by default.

Seven current captures from an isolated local dev ring: the quoting flow from vehicle spec to a real-cost estimate and term comparison, Claudio's guided search, and a flag-gated partner-application intake. All amounts, applicant and contact details are synthetic, and every estimate is illustrative, not a lessor offer. Synthetic demo data, captured 2026-09-09.

What's working, and what's still in progress

Shared domain rules and recoverable document handoffs keep operating decisions consistent.

What's working

  • A quoting flow that turns a vehicle spec into a real-cost estimate — rent, estimated ISR deduction, creditable IVA — compared across terms
  • Claudio, a guided-search adviser that carries a confirmed budget into vehicle questions
  • A partner-application intake, gated behind a feature flag rather than live

Still in progress

  • Every amount is a local illustrative estimate, not a lessor offer — the product itself labels it "Estimación, no oferta"
  • Claudio runs on a deterministic scripted mock (USE_REAL_AGENT_MODEL off); no live model call and no autonomous decision
  • The partner application sits behind USE_PARTNER_PORTAL_V2, off by default while the partner program stays intentionally dark
  • The vehicle catalog itself isn't shown here: its cards currently render an "Imagen en validación" placeholder because licensed vehicle media doesn't exist yet

One goal, many systems

The customer's goal is simple: find a vehicle and understand whether leasing it makes sense. Behind it sit vehicle data, pricing and tax rules, eligibility, documents, partners, outside systems and human review.

Improving one screen would not fix that. Quotes have to apply the same rules everywhere, prequalification has to keep its context, and operators need to see what happened, what failed and what to do next.

What I do

I set technical direction and stay hands-on in the implementation, across product, engineering, operations and outside constraints:

  • Product architecture and delivery sequencing
  • Customer web and mobile experiences
  • Vehicle discovery, quoting and prequalification
  • Conversational intake and bounded applied-AI workflows
  • Partner, operator, document and contracting handoffs
  • Shared domain models, integrations, observability and release gates
Discovery, quoting, prequalification, routing and operations share one rule authority, and failures return to an owned recovery lane.
A sanitized handoff map: one rule authority, recoverable handoffs.

Decisions

  • Keep quote and eligibility rules in one place instead of copying them across customer, partner and operator surfaces.
  • Carry context through multi-step journeys, so coming back doesn't restart the process or lose earlier answers.
  • Model each handoff as an owned state with explicit inputs, outputs and recovery.
  • Keep models inside bounded workflows, never as silent business authority.
  • Treat timeouts and partial failures as normal operating states, not afterthoughts.
  • Keep source, CI, runtime and production evidence separate, and tie every release claim to the artifact that was actually verified.

Evidence boundary

This covers my ongoing work on the new PAZZ platform, which is in development and not publicly deployed. The current company website is not my work and is not a demo of this platform.

It leaves out customer data, supplier identities, private architecture, incidents, security details and business metrics, and it claims no revenue, conversion or approval outcomes. The screens above come from an isolated development build with synthetic data.

Stack

TypeScript, Next.js, React Native with Expo, PostgreSQL on Supabase, GitHub Actions, Sentry, and applied AI inside bounded workflows.