Skip to content

Backends for production AI

I design typed services, agent tools and cloud workflows that move AI from demos into regulated products, currently for banking and fintech teams.

See the systemResume, opens in a new tab
Portrait of Oscar Bucio in a suit
Oscar BucioCDMX
Status
Open to roles
Work auth
U.S. citizen · No sponsorship
Latest
Agent tools for BBVA
Revision
Rev. 2026.09
API requests a year
60M+Serverless AWS platform I helped build and operate at Siatech.
Users served
12,000+Same platform, kept under a 0.3% Lambda error rate.
Cost per request
−39%From MongoDB query, batching and caching work.
Automated tests
7,000+Behind CI/CD with GitHub Actions across 20+ services.

Anatomy of a production agent

The architecture I build toward. Requests enter through an authenticated edge, an orchestrator works through typed tools, grounded retrieval and model calls, and nothing reaches a user until three independent checks agree.

Fig. 02 · Request lifecycleLive simulation
Production AI agent architectureRequests from WhatsApp, a web app and internal services pass through an authenticated edge to an orchestrator. The orchestrator calls typed tools, retrieval and models, then sends its draft to a gate where policy, grounding and schema checks must all approve before the reply is sent. Every step reports to an observability bus, and state is written to an event bus.OBSERVABILITYtraces · evals · error trackingSTATE & EVENTSMongoDB sessions · SQS · EventBridge · Step FunctionsWHATSAPPwebhooksWEB APPREST · SSEINTERNALservice APIsEDGEgateway · authORCHESTRATORagent runtimeTOOLStyped Python · PydanticRETRIEVALembed · search · rerankMODELSBedrock · OpenAIGATE3 / 3 to shipREPLYto userPLANACTCHECK≤ 6 STEPS · BUDGETEDPOLICYGROUNDSCHEMA

Swipe to explore

Request Active path Telemetry and stateValidation core
  1. 01

    Edge

    Webhooks and APIs end at an authenticated gateway. Tenancy is resolved before a model sees a single token.

  2. 02

    Orchestrator

    An agent runtime plans, calls tools and drafts an answer inside explicit step and cost budgets.

  3. 03

    Capabilities

    Typed Python tools with Pydantic contracts, retrieval with reranking, and a model layer with provider fallback.

  4. 04

    Gate

    Policy, grounding and schema checks vote independently. A draft ships only on three of three.

  5. 05

    State and telemetry

    Sessions and events persist to MongoDB and queues, and every hop emits traces and eval signals.

Systems shown in context

Banking agents, a high-scale serverless platform and financial retrieval, with the boundaries and results that shaped each build.

Case 01Agent Blue, global core team

Plexus Tech, client BBVA

Tool modules for a global banking agent.

Built production tool modules used across global and Mexico-specific agent workflows, plus the regional adaptation patterns that let Mexico adopt the global tools module instead of duplicating it.

  • Python
  • OpenAI Agents SDK
  • Pydantic
  • Async service clients
  • Enterprise auth
  • End-to-end tests
Global agent, regional tools
  1. Agent orchestrationOpenAI Agents SDK
  2. Global tools moduleTyped Python, Pydantic contracts
  3. Regional adaptationMexico-specific workflows
  4. Banking servicesAsync service clients

Case 02Serverless platform and LLM agent

Siatech

A serverless platform at 60M requests a year.

Helped build and operate a serverless AWS platform for 12,000+ users, including an LLM-powered WhatsApp agent that handles 140,000+ webhooks a month. Cut cost per request 39% through MongoDB query, batching and caching work.

  • Python
  • AWS Lambda
  • MongoDB
  • LLM agents
  • WhatsApp webhooks
  • GitHub Actions
WhatsApp LLM agent, monthly
  1. 140K+webhooks
  2. 30K+LLM turns
  3. 1,000+leads

Over 90% processing success. CI/CD with GitHub Actions across 20+ services, backed by 7,000+ automated tests.

Case 03RAG conversational agent

FIXAT

Financial RAG, from prototype to production.

Owned the design, implementation and production deployment of a RAG agent used by thousands of customers, improving lead-capture efficiency by up to 60%.

  • Amazon Bedrock
  • LangChain
  • OpenAI API
  • AWS Lambda
  • Step Functions
  • SQS
  • EventBridge
  • MongoDB
Retrieval and event orchestration

Retrieval runtime

Amazon Bedrock + LangChain

Serverless compute

AWS Lambda

Workflow orchestration

Step Functions

Messaging

SQS

Event routing

EventBridge

Persistence

MongoDB

Engineering that holds under pressure

I keep AI systems grounded in explicit interfaces, controlled access and workflows that teams can operate.

  1. 01

    Contracts before cleverness

    Typed boundaries keep business rules, agent tools and infrastructure understandable as the system grows.

    Pydantic · typed Python

  2. 02

    Agents use explicit tools

    Models can reason, but production services still own policy, validation, authentication and data access.

    Tool contracts · auth

  3. 03

    Operability is part of design

    Clear failure paths, observable workflows and testable components matter before a feature reaches production.

    Traces · evals · tests

From automation to production AI

Each role moved closer to owning the services behind intelligent products.

Full history on LinkedIn, opens in a new tab

  1. Feb 2026 – Sep 2026

    Plexus Tech, client BBVA

    Global banking agent tools

    Full-Stack AI Engineer
    Mexico City

  2. Jan 2025 – Dec 2025

    Siatech

    Serverless platform and LLM agent

    Software Engineer, Backend, AI & Cloud
    Remote

  3. Jul 2023 – Dec 2024

    FIXAT

    Production RAG agent

    Software Engineer
    Remote

  4. Jan 2022 – Jun 2023

    BASF Mexicana

    Workflow automation and reporting

    Python Automation / Digitalization Intern
    Mexico City

B.S. Computer Systems Engineering · Instituto Politécnico Nacional (ESCOM) · 2025

Let’s build dependable AI

Open to backend and AI engineering roles with U.S. teams. U.S. citizen based in Mexico City, no sponsorship required.

Email me

Channels