dderas.dev

Software Engineer · San Salvador, El Salvador

I build AI agents and RAG systems that ship.

Agent harnesses, retrieval pipelines and the full-stack platforms that run them, for public-sector services. Currently pursuing a Machine Learning specialization.

Get in touch LinkedIn GitHub

This is a real similarity search over embeddings of my work.

01

What I build

· the camera travels to each cluster as you scroll

Agents

Agent harnesses

Runtimes with administrable skills, tool registries, lifecycle hooks and MCP clients, with queued executions, guardrails and human handoff.

  • MCP
  • Tool calling
  • Claude · Gemini
  • Guardrails

RAG

Retrieval systems

Layout-aware chunking with Docling, embeddings indexed with HNSW in pgvector, and hybrid semantic and full-text search.

  • Docling
  • pgvector
  • Hybrid search
  • FastAPI

ML

Applied ML

Computer vision with honest evaluation: source-grouped splits, fairness audits and int8 models running offline on a phone.

  • PyTorch
  • TFLite int8
  • Optuna
  • OOD detection

Full stack

Platforms

Laravel, Node.js and Vue 3 services on Google Cloud, built for thousands of interactions across web, Telegram and WhatsApp.

  • Laravel
  • Vue 3
  • Cloud Run
  • Pub/Sub
02

Selected work

DoctorMaiz

Edge AI / Computer vision

0.947

macro-F1 · 3.56 MB · ~60 ms offline on Android

Case study

FIA Incident Management System - UES

Incident management / Maintenance

6,000+

students and 200+ staff served

Case study

Vynta

Productivity tool / Desktop

MS Store

published desktop app

Case study
All work
03

Experience

  1. 2024 — now

    Software Engineer (Full Stack & AI)

    Secretaría de Innovación de la Presidencia

    • Built LLM agent-harness components (runtime, skills, tool registry, lifecycle hooks, MCP client, queued executions) for a multi-tenant platform designed to serve 200+ public institutions.
    • Built RAG pipelines: Docling-based document chunking in a FastAPI microservice, Gemini embeddings indexed with HNSW in PostgreSQL/pgvector, and hybrid semantic and full-text search.
    • Developed tool-calling assistants on a provider-agnostic LLM layer (Claude, Gemini, OpenAI-compatible) with automatic fallback, guardrails and per-turn token-cost observability.
    • Built Laravel and Node.js microservices, REST APIs and Vue.js/Vuetify frontends for thousands of interactions across web, Telegram and WhatsApp.
  2. 2023

    Frontend Developer Intern

    Secretaría de Innovación de la Presidencia

    • Built Vue.js and Vuetify interfaces integrated with REST APIs for institutional web systems.
  3. 2023

    Full Stack Developer

    Freelance (Retail & Warehouse Client)

    • Built a custom ERP for inventory, HR, finance and administrative workflows using Scrum.
04

Writing

Probability and Machine LearningProbability allows us to model uncertainty and make informed decisions. In this article, we will explore how this tool is used in the context of machine learning.machine learningStatistics and Machine LearningStatistics allows us to analyze and understand data. In this article, we will explore more about descriptive and inferential statistics, and how these concepts are fundamental for machine learning.machine learningVectors and Machine LearningVectors are a fundamental part of machine learning, as they allow for the efficient representation of data and model parameters. In this article, we will explore their definition and application in machine learning algorithms.machine learning
All articles

Let's build something useful.

Open to roles building LLM agents, RAG systems and applied machine learning.

david@deras.dev LinkedIn GitHub