Gilberto Carrillo

Gilberto Carrillo

AI Engineer & Team Lead

Venezuela · Open to remote · IEEE-published researcher

Building AI agentic systems and the architecture behind them.

AI Engineer & Team Lead — 3 years building production LLM platforms, workflow engines and agents in Python, FastAPI and Azure.

Gilberto Carrillo

Gilberto Carrillo

AI Engineer & Team Lead

Agentic AI · LLM Platforms · Python Backend

Venezuela · Open to remote · IEEE-published researcher

About me

I'm an AI engineer who builds the backend behind agentic products: the infrastructure that turns "wire an LLM into a feature" into something reliable enough for real clients.

At MAIDEPOT, a South Korea-based platform for AI workflows and agents, I started as a Backend Engineer and was promoted to Team Lead within six months, while staying hands-on as the platform's main backend contributor. I built much of its AI core: a model-agnostic layer across 7 LLM providers, a multi-model chat, a DAG workflow engine and the agents built on top of it.

Now

Open to remote AI engineering roles building agentic products and LLM platforms.

Currently deepening my cloud and infrastructure skills (Azure, Docker, Kubernetes).

Based in
Venezuela, working remotely
Focus
Agentic AI, LLM platforms, Python backend
Languages
Spanish (native), English (professional working proficiency)
Education
B.Sc. Systems Engineering, Universidad de los Andes
Certifications
  • Neural Networks and Deep LearningDeepLearning.AI
  • Machine Learning SpecializationDeepLearning.AI & Stanford Online

Systems I've built

Selected work at MAIDEPOT ↗, a South Korea-based platform for AI workflows & agents.

Model layer

One interface, seven LLM providers

Built a model-agnostic LLM layer across OpenAI, Azure OpenAI, Anthropic, Gemini, DeepSeek, GLM and LG EXAONE, with retries, rate-limit failover and load balancing, so features could switch models through configuration. On top of it, built the platform's multi-model chat: SSE streaming, file attachments checked against each model's capabilities, web-search citations and per-message credit accounting. It later became the foundation for the agent experience.

Python · FastAPI · SSE · Multi-provider LLMs


Workflow engine

From a linear prompt queue to a DAG engine clients build on

Re-architected the workflow engine from a linear prompt queue into a JSON-defined DAG with parallel execution, LLM-evaluated conditional branches and map-reduce loops over documents, images, audio and video. Enterprise clients such as KOTRA and The Korea Times built their own workflows on it. It also powers two of the platform's most-used features: translation of documents up to 1,000 pages and transcription of long audio and video, both consistent across chunks and resumable after failures.

Python · FastAPI · DAG · Map-reduce · OCR · Chunking


Agents

Agents that use the whole platform as tools

Extended the platform to PydanticAI agents: workflows became callable tools alongside ~40 native tools for files, web search, image and video generation and knowledge access, plus scheduled routines with cost limits. Also designed and built semantic tool discovery for large tool libraries, ranking tools in Elasticsearch by relevance and real usage. Delivered client agents such as a sales-training assistant for LG.

PydanticAI · Tool calling · Elasticsearch · Embeddings


Knowledge

RAG knowledge bases and an LLM wiki

Built RAG for agent knowledge bases, with structure-aware chunks, pgvector and Cohere rerank, and a vector-free "LLM wiki" engine in which agents navigate an index of pages and structured tables. Also built the platform's document drive, where files, archived chats and workflow outputs are stored, summarized and reused as input in chats, agents and workflows.

RAG · pgvector · Cohere Rerank · PostgreSQL


Experience

  1. Backend Engineer & Team Lead — AI Platform

    Jan 2024 – Aug 2026

    Leadership

    • Promoted to Team Lead within six months; led a cross-functional engineering team (frontend, full-stack and QA) while remaining the main backend contributor and owning architecture decisions.

    AI platform

    • Built a model-agnostic LLM layer across 7 providers, with retries, rate-limit failover and load balancing, and the multi-model chat on top of it.
    • Re-architected the workflow engine into a JSON-defined DAG with parallel execution, LLM-evaluated branches and map-reduce loops; KOTRA and The Korea Times built their own workflows on it.
    • Delivered translation of documents up to 1,000 pages and transcription of long audio and video, two of the platform's most-used features.

    Agents & knowledge

    • Extended the platform to PydanticAI agents with ~40 native tools, workflows as tools and scheduled routines; delivered client agents such as a sales-training assistant for LG and Spanish tutors for Korean language academies.
    • Built RAG knowledge bases (pgvector, Cohere rerank) and a vector-free LLM wiki engine.

    Backend, API & operations

    • Owned the core FastAPI backend: SSE streaming, business-level multi-tenancy, Redis locks and rate limiting.
    • Built the public API (hashed API tokens, rate limits per operation), usage-based billing priced per model and token type, the document drive and the enterprise dashboard.
    • Managed releases on Azure App Service across dev, staging and production with zero-downtime slot swaps; ran heavy workloads on Azure Functions with Storage Queues.
  2. Promoted · Jan 2024
  3. Backend Engineer

    Jul 2023 – Dec 2023

    • Migrated the platform's core backend, including its original text and image generation apps, from Google Cloud Functions to a dedicated FastAPI service, eliminating cold starts and consolidating the logic into a single, maintainable API.

Trusted in production

  • LG Sales-training assistant agent.
  • The Korea Times Article generation on the workflow engine.
  • KOTRA Report generation and translation.
  • Language academies Spanish-learning tutor agents, South Korea.

Research

Before building AI products, I worked on AI for learning: estimating student skill with reinforcement learning, and teaching with robots and speech recognition.

  1. IEEE paperLACLO · 2022

    Reinforcement Learning for Estimating Student Proficiency in Math Word Problems

    Using reinforcement learning to estimate how well a student solves math word problems.

    Perez, J., Dapena, E., Aguilar, J., & Carrillo, G. Proceedings, pp. 1–6.

    Indexed in IEEE Xplore · 2022 XVII Latin American Conference on Learning Technologies (LACLO)

    Read on IEEE
  2. ThesisUniversidad de los Andes · 2022

    Design of Human-Robot Interaction for Learning Multiplication Tables and Implementation of Speech Recognition Models

    A robot that helps children practice multiplication tables, with speech recognition so they can answer out loud.

    Carrillo, G. B.Sc. in Systems Engineering, Universidad de los Andes.

    2nd place · VII National Contest of Special Degree Projects, Venezuelan Computing Society. View certificate ↗

    Read thesis

Toolkit

The stack, top to bottom.

  1. Interface

    • Next.js
    • React
    • JavaScript
    • Redux
    • Vercel
  2. Agents

    • PydanticAI
    • MCP
    • Workflow orchestration
    • Tool use / function calling
  3. Models & retrieval

    • Multi-provider LLM orchestration
    • RAG
    • pgvector
    • Embeddings & reranking
    • Prompt engineering
    • Context management
    • Azure Document Intelligence
    • AssemblyAI
  4. Backend

    • Python
    • FastAPI
    • PostgreSQL
    • Redis
    • Elasticsearch
  5. Cloud & infra

    • Azure (App Service, Functions, OpenAI)
    • Firebase
    • Docker
    • Git/GitHub

Building something with agents? Let's talk.

Open to remote AI engineering roles building agentic products and LLM platforms.