ES

› requestI need AI working inside my business, under control.

OsmarLiera

Applied AI EngineerBackend / Full-Stack

I design and ship LLM agents with tool calling, structured outputs and human-in-the-loop approval, integrated with business data and APIs.

  • La Paz, Baja California Sur, Mexico
  • Open to remote or hybrid roles
See AI case studiesView résuméWhatsApp(opens in a new tab)
Close-up portrait of Osmar Liera in a blue polo shirt
  1. request“Who should I follow up with today?”
  2. agentplans the day from the follow-up queue
  3. toolfollow-up-queue
  4. tooldraft-whatsapp-message
  5. approvalsend-whatsapp-message · awaiting human approval
  6. resultapproved, sent and written to the audit log
Illustrative example of an assistant run on the OS CRM MCP server.

01agent

Software that puts AI to work inside real businesses.

I am a software engineer with 4 years of professional experience building, integrating and running business applications in production for warehousing, retail, insurance and healthcare.

Today I work on four fronts: at Nuts Marketing I build integration middleware between Shopify stores and Microsoft Dynamics 365 ERPs and look after the Shasa (Mexico and US) and Hypsters platforms; at Tangramx I work in Python (FastAPI), where I shipped LLM-based OCR to production and evaluated it against ground truth before rollout; at ELROI Labs, as founder and technical lead, I built the AI layer of our CRM; and at the Instituto Tecnológico de La Paz I program the institutional systems and teach the final semester of the software development programs.

My development workflow is AI-assisted end to end with Claude Code: implementation, tests, refactoring and code review.

Where I work today

  • Nuts Marketing

    Present

    Full-Stack Developer

    Middleware between Shopify stores and Microsoft Dynamics 365 ERPs, plus the Shasa (Mexico and US) and Hypsters platforms.

  • Tangramx

    Present

    Full-Stack Developer

    Management applications for insurers, healthcare institutions and logistics operators. Primary stack: Python / FastAPI.

  • ELROI Labs

    Present

    Founder & Technical Lead · La Paz, BCS

    Software, integration and applied-AI studio. I lead technical operations and deliver projects end to end.

  • Instituto Tecnológico de La Paz

    Present

    Institutional Systems Programmer · Teacher

    Programmer of the institutional systems and teacher in the final semester of the software development programs.

Languages

Spanish
Native
English
Intermediate (B1–B2): daily technical reading and written communication

02tools

AI in production: case studies

Four real systems, described as they work: permissions, audit trails, validated outputs, provider fallback and evaluation before rollout.

ELROI Labs · OS CRM (Laravel, React)

An MCP server that lets AI assistants operate the CRM

The agent layer of our internal CRM for leads, pipeline, multichannel follow-up, clients and projects.

  • Built an MCP server with 23 tools, 3 prompts and a playbook resource so AI assistants (Claude and others) can search leads, plan the day, log follow-ups, schedule tasks, import leads and generate pipeline and activity reports.
  • Agents run with the token owner’s permissions and every change is written to the audit log.
  • Outbound WhatsApp messages require explicit human approval before sending.
  • Covered the MCP server, the AI assistant and the WhatsApp flows with automated feature tests.
  • MCP
  • Tool calling
  • Human-in-the-loop
  • Audit log
  • Laravel
Architecture of the OS CRM MCP serverAn AI assistant calls the MCP server (23 tools, 3 prompts and one resource). Every call runs under the token owner’s permissions and is written to the audit log before it touches the CRM. Outbound WhatsApp messages stop until a human approves them.

  1. client

    • Claude or another assistantMCP client

  2. MCP server

    • 23 toolsleads, follow-ups, tasks, reports
    • 3 prompts + 1 resourcefollow-up playbook

  3. control

    • Token owner’s permissions
    • Audit log

  4. effect

    • CRMleads · pipeline · tasks
    • Outbound WhatsApprequires human approval

ELROI Labs · OS CRM

LLM agents that answer in JSON Schema, not free text

Agents with structured outputs the system can validate and use directly.

  • A lead-contact assistant that drafts messages, call scripts and next tasks.
  • A project planner that breaks scope into tasks with acceptance criteria and estimates.
  • A WhatsApp writer grounded in the real conversation and an editable business context.
  • Guardrails against hallucination: answer only from supplied data, ask instead of assuming, no prices or commitments; per-agent temperature and token limits.
  • Structured outputs
  • JSON Schema
  • Guardrails
  • Prompt design
Flow of the structured-output agentsThe agent only receives supplied context (lead data, the real conversation and an editable business context), applies guardrails and returns a structured output validated with JSON Schema that becomes messages, call scripts, tasks or WhatsApp drafts.

  1. context

    • Lead data
    • Real conversation
    • Business contexteditable

  2. agent

    • LLM agentper-agent temperature and tokens
    • Guardrailssupplied data only

  3. output

    • JSON Schemastructured, validated output

  4. use

    • Messages, scripts and tasks
    • Tasks with criteria and estimates
    • WhatsApp draft

ELROI Labs · OS CRM

An OpenAI-compatible model gateway with fallback

One interface so the agents never depend on a single provider.

  • Wrote a custom OpenAI-compatible model gateway with model-catalog discovery, caching and fallback.
  • Multi-provider setup: OpenAI, Anthropic, Gemini, Azure and others.
  • OpenAI-compatible
  • Fallback
  • Caching
  • Multi-provider
OS CRM model gatewayThe OS CRM agents call an OpenAI-compatible API. The gateway discovers and caches the model catalog and, if a provider fails, falls back to another one: OpenAI, Anthropic, Gemini, Azure or others.

  1. app

    • OS CRM agents

  2. gateway

    • OpenAI-compatible API
    • Model catalogdiscovery and caching
    • Fallbackacross providers

  3. providers

    • OpenAI
    • Anthropic
    • Gemini
    • Azure and others

Tangramx · Vitase / Paciente Integral

LLM-based OCR, evaluated against ground truth before rollout

Medical records and claims management for insurers (Jalisco Health Services).

  • Integrated OpenAI Vision for OCR of Mexican national ID cards (INE), replacing the previous engine.
  • Did not trust the model blindly: built a deterministic validation layer on top of its output for CURP, with check-digit reconstruction, homoclave recovery and edge cases (Ñ, MARÍA/JOSÉ compound names).
  • Evaluated against ground truth: 25 real ID cards with 100% name accuracy; documented the data-quality findings that guided rollout.
  • OpenAI Vision
  • Evaluation
  • Deterministic validation
  • Python / FastAPI
INE ID-card OCR pipelineThe ID-card image goes through OpenAI Vision. A deterministic layer validates the CURP (check digit, homoclave and edge cases). The output was evaluated on 25 real ID cards, with 100% name accuracy, before rollout.

  1. input

    • INE ID cardimage

  2. model

    • OpenAI VisionLLM-based OCR

  3. validation

    • CURPcheck digit and homoclave
    • Edge casesÑ, MARÍA / JOSÉ

  4. evaluation

    • 25 real ID cards100% name accuracy
    • Rolloutguided by data quality

03human approval

Human control and evidence over magic.

How I take an agent from demo to production without losing control.

  1. User permissions

    Agents act with the token owner’s permissions, never more.

  2. Audit log

    Every change an agent makes is recorded.

  3. Explicit approval

    No message reaches a customer until a person approves it.

  4. Guardrails

    Supplied data only, ask instead of assuming, and no prices or commitments.

  5. Evaluation

    Model output is measured against ground truth before rollout.

  6. Automated tests

    MCP server, AI assistant and WhatsApp flows covered by tests.

04result

Experience: production systems for real businesses.

Four active positions. No fluff: what I built, what I fixed and how I validated it.

  1. 2024 — Present

    Nuts Marketing

    Full-Stack Developer

    Integration middleware between Shopify stores and Microsoft Dynamics 365 ERPs, plus the Shasa (Mexico and US) and Hypsters platforms. Stack: Laravel, Laravel Nova, Filament, Shopify Admin API (GraphQL and webhooks), SOAP and Azure App Service with WebJobs.

    • Macadamia (internal project) — Shopify ↔ Dynamics 365 Finance & Operations

      Middleware for a retailer running Dynamics 365.

      • Order and customer sync, store pickup by location and payment-method mapping.
      • Reconciliation reports across Shopify ↔ middleware ↔ Finance & Operations.

    • El Asturiano — Shopify ↔ Business Central + LS Central

      Integration with Dynamics 365 Business Central and LS Central over SOAP/XML.

      • Receives Shopify webhooks and sends the data to the ERP over SOAP/XML.
      • Inventory sync with scheduled jobs.

    • Maratón

      Digital development for the classic Maratón trivia game.

    • Coco (internal project)

      New Laravel + Nova application on Azure App Service, getting started.

      • Bootstrapping the application with Laravel and Laravel Nova, deployed on Azure App Service.

    • Shasa (Mexico and US) and Hypsters

      Fashion retail with large catalogs in two markets.

      • Backend development and Shopify and Magento integrations syncing catalogs, inventory, orders and customers.
      • Apple Wallet passes and gift cards for Shasa.
  2. Present

    ELROI Labs

    Founder & Technical Lead

    La Paz, BCS

    Software, integration and applied-AI studio. I lead technical operations and deliver projects from design to deployment and production support.

    • OS CRM — in production

      CRM for leads, pipeline, multichannel follow-up, clients and projects (Laravel, React).

      • Designed, deployed and maintain the platform in production, built with an AI-assisted workflow (Claude Code) and covered by automated feature tests.
      • WhatsApp integration for sending text and media from the CRM, logged to each lead’s history; queued notifications and task reminders.
      • AI layer: an MCP server with 23 tools, LLM agents with structured outputs and an OpenAI-compatible model gateway with fallback.
      See the AI case studies

    • Client work

      Applied AI.

      • Applied AI for customer service, lead prequalification and follow-up automation.
  3. Present

    Tangramx

    Full-Stack Developer

    Enterprise applications for logistics operators, insurers and healthcare institutions. Primary stack: Python / FastAPI.

    • Vitase / Paciente Integral — LLM-based document OCR

      Medical records and claims management for insurers (Jalisco Health Services).

      • Integrated OpenAI Vision for OCR of Mexican national ID cards (INE), replacing the previous engine.
      • Deterministic CURP validation layer and evaluation on 25 real ID cards: 100% name accuracy.
      Read the full case

    • Compass WMS — warehouse management system

      Warehouse operations and client portal (internal e-commerce).

      • Built the Administrator preview mode for the Client Portal and fixed folio assignment in the order flow.
      • Removed a hidden pagination ceiling by implementing true server-side pagination for large datasets.
      • Implemented the product gallery with image reordering, fixed a race condition in multi-file upload and resolved session and authentication bugs.

    • ATC · ATCSMART

      Medical account-statement processing and insurance-claims systems (FastAPI).

      • Fixed incorrect discount detection and an off-by-one date shift in the statement parser.
      • Added Excel exports segmented by insurer and hospital and wrote the local setup guides.
  4. Present

    Instituto Tecnológico de La Paz

    Institutional Systems Programmer · Teacher

    A Tec graduate with a B.S. in Computer Systems Engineering (Software Development specialization); today I work there in two roles: building its institutional systems and teaching.

    • Institutional systems programmer

      Development of the Tec’s institutional systems.

    • Teacher

      Final semester of the software development programs.

      • As part of teaching, training on applied AI for business: process automation, operational productivity and AI tool adoption, plus AI for social media and content.

Previous experience

  • Giro26

    Full-Stack Developer · 2023 — 2024

    Financial ERP, hexagonal architecture and SQL optimization.

  • Hi-G

    Web & Mobile Developer · 2024 — 2025

    Hi-G, Karol G’s app: Flutter, Firebase, TypeScript and Web3 integrations.

E-commerce across two markets

Backend work and integrations on Shopify and Magento for fashion retail with large catalogs in Mexico and the US: catalogs, inventory, orders and customers synced with internal systems.

Work delivered as part of the Nuts Marketing team. Shasa retains brand and product ownership.

Own products

Four products of my own. Each one solves a different operation.

  1. CRM with an AI layer

    OS CRM

    ELROI Labs’ CRM, in production: leads, pipeline, tasks with reminders, clients and projects.

    • Built-in WhatsApp (Evolution API) with AI-written drafts and human approval before sending
    • MCP server with 23 tools so AI agents can operate it, with permissions and auditing
    • Postiz webhook to record social media publications, and website forms that create leads

    Deployed on Coolify

    • Laravel 13
    • Inertia
    • React
    • Evolution API
    • Laravel MCP
    • Coolify

    Private · in production

    OS CRM mapLeads come in through website forms, the Postiz webhook records social media publications and WhatsApp arrives through Evolution API. OS CRM manages leads, pipeline, tasks with reminders, clients and projects. AI drafts wait for human approval before sending, and agents operate the CRM through an MCP server with 23 tools, permissions and auditing.

    1. inputs

      • Website formscreate leads
      • Postiz webhooksocial publications
      • WhatsAppEvolution API

    2. os crm

      • Leads and pipeline
      • Taskswith reminders
      • Clients and projects

    3. control

      • AI draftshuman approval before sending
      • Permissions and audit log

    4. agents

      • MCP server23 tools

Stack, with evidence

Every technology is tied to the project where I used it. If no real project backs it, it is not listed here.

AI & agents

AI & agents
TechnologyEvidence
MCP servers (tools, prompts, resources)
  • OS CRM
Tool calling and human approval
  • OS CRM
Structured outputs with JSON Schema
  • OS CRM agents
OpenAI-compatible gateway · multi-provider
  • OS CRM
Laravel AI SDK · Laravel MCP
  • OS CRM
OpenAI Vision
  • INE OCR · Tangramx
Ground-truth evaluation
  • INE OCR · Tangramx

Backend

Backend
TechnologyEvidence
Python · FastAPI
  • Tangramx: Vitase, ATC, Compass WMS
SQLAlchemy · Alembic
  • Tangramx
Laravel (PHP)
  • Nuts Marketing
  • OS CRM
  • PMCRM
  • AVT
  • NopalGreen
Laravel Nova · Filament
  • Nuts Marketing: Coco and middleware
REST APIs
  • Tangramx (FastAPI)
  • AVT (Sanctum API)
Queues, scheduled jobs and webhooks
  • Nuts Marketing: Shopify webhooks and inventory
  • OS CRM: notifications and reminders
Authentication with 2FA
  • PMCRM
  • NopalGreen

Integrations

Integrations
TechnologyEvidence
Shopify Admin API (GraphQL and webhooks)
  • Nuts Marketing: Shasa, Hypsters, El Asturiano, Macadamia
Microsoft Dynamics 365 (Finance & Operations, Business Central + LS Central)
  • Nuts Marketing: Macadamia, El Asturiano
SOAP / XML
  • Nuts Marketing: El Asturiano
Magento
  • Nuts Marketing: Shasa, Hypsters
Apple Wallet (passes and gift cards)
  • Nuts Marketing: Shasa
WhatsApp via Evolution API
  • OS CRM
Postiz webhooks and website forms
  • OS CRM
Google Maps
  • AVT

Data

Data
TechnologyEvidence
SQL and relational modeling
  • OS CRM
  • AVT
  • NopalGreen
Versioned migrations
  • Alembic · Tangramx
  • Laravel · OS CRM
Spatial data
  • AVT
Excel exports
  • ATCSMART
  • AVT

Frontend

Frontend
TechnologyEvidence
React · TypeScript
  • OS CRM
  • PMCRM
  • NopalGreen
Inertia
  • OS CRM
  • PMCRM
  • NopalGreen
Tailwind CSS · Vite
  • OS CRM
  • PMCRM
  • AVT
  • NopalGreen
Astro
  • This portfolio

Cloud & ops

Cloud & ops
TechnologyEvidence
Azure App Service and WebJobs
  • Nuts Marketing: in production, Coco
Coolify · containers · Linux
  • OS CRM in production
CI with GitHub Actions
  • OS CRM: tests and static analysis
Automated tests (Pest)
  • OS CRM
Claude Code · Git/GitHub branching, PRs and code review
  • OS CRM
  • Daily development

Open to remote or hybrid roles

Does your team need to ship AI to production, under control?

I am looking for Applied AI Engineer, Backend or Full-Stack roles, remote or hybrid. Send me a message.

Message on WhatsApp(opens in a new tab)Send an email
Phone & WhatsApp
+52 615 155 9659