Alejandro Zanello

AI Engineer — Agent Runtimes, Multi-Agent Systems & GenAI

[email protected] | +54 9 351 214 1761 | Córdoba, Argentina

azanello.com | linkedin.com/in/alejandrozanello | github.com/AleSZanello

Languages: Spanish (Native) · English (Full professional)

Professional Summary

AI Engineer with 6+ years shipping production software across mobile, backend, and web, now focused on agent runtimes and multi-agent systems. Author of Akashi, an open-source agent framework built as a 14-package monorepo with 13 packages published on pub.dev: a streaming tool-calling runtime, token-accurate context management, multi-provider failover by error type, durable checkpoint/resume with human-in-the-loop, OpenTelemetry tracing, and structured-output self-repair to contain non-deterministic model behaviour. Builds LLM systems with LangChain, LangGraph, RAG, and the Model Context Protocol (MCP) in TypeScript and Python. Certified: IBM Generative AI Engineering Professional Certificate, Agentic AI Engineering Specialization and AWS Certified Cloud Practitioner.

Skills

AI & Agentic: LangChain, LangGraph, LangSmith, MCP, RAG, Evals, Langfuse, pgvector, Anthropic SDK, OpenAI, Google Gemini, Vercel AI SDK, Vertex AI, Claude Code, Codex, Cursor, Windsurf, GitHub Copilot

Languages: TypeScript, PHP, Python, Dart, Rust, Swift

Frontend: React, Next.js, Tailwind CSS, Svelte

Backend: Node.js, NestJS, GraphQL, Prisma, FastAPI, Hono, OAuth 2.1, MercadoPago

Mobile: Flutter, React Native, Ionic, RevenueCat, Codemagic, fastlane

ML & Data: PyTorch, Hugging Face, scikit-learn, Reinforcement Learning, DuckDB, Parquet

BaaS & Databases: Supabase, Firebase, PostgreSQL, MySQL

Low-Code & Automation: FlutterFlow, GitHub Actions, Make, n8n

DevOps: Docker, Git, GCP, AWS, Vercel, Cloudflare Workers, Linux (systemd)

Testing: Playwright, Maestro, Selenium

GenAI & ML (certified): LLMs, Transformers, Fine-Tuning, Agentic AI

Projects

Akashi

2026

Dart · Flutter · Multi-Agent · MCP · RAG · OpenTelemetry

Provider-neutral agent framework for Dart & Flutter — a 14-package monorepo with 13 packages published on pub.dev, bringing the agent runtime primitives Python and TypeScript take for granted to the Dart ecosystem. Multi-agent orchestration via subagent-as-tool, control handoffs, and per-step model escalation.

  • Agent runtime: streaming-first tool loop with typed tools and dependency injection, composable stop conditions, and a prepareStep context hook — last-N windowing plus summarize-older-than compaction via a cheaper model.
  • Multi-provider resilience: FallbackModel fails over across Gemini, OpenAI, and Anthropic behind one contract, taking a shouldFailover(error) predicate to fail over by error type, and blocking mid-stream failover because re-running an emitted stream duplicates output.
  • Durable execution: versioned Message/Part serialization over a SQLite checkpoint store; human-in-the-loop approvals suspend a run and resume it in another process.
  • Non-determinism hardening: structured output with self-repair; three providers' streaming formats (text, thinking, tool_use) normalized into one sealed union with exhaustive matching.
  • Operability: OpenTelemetry tracing, per-call token usage accounting, MCP tools, and bounded-concurrency workflows with retries, backoff, timeouts, and cancellation.

mcp-deep-dive

2026

TypeScript · MCP · pnpm

Protocol-first MCP reference in TypeScript: custom server + client + all 3 capabilities (tools, resources, prompts), both transports (stdio + HTTP), with wire-level JSON-RPC logging. No LangChain wrappers.

mcp-standard

2026

MCP · OAuth 2.1 · Security · Architecture

MCP Server Reference Architecture: a language- and framework-agnostic standard for exposing a product to AI agents through the Model Context Protocol. Fixes layers, contracts and invariants — role-scoped reads, OAuth 2.1 and API keys behind one gateway, propose → apply writes with a human yes, prompt-injection defense, and an audit row for every call — with three conformance levels and annexes for implementers and coding agents.

caucionetta

2026

Python · Swift · MCP · SwiftBar · macOS · launchd

Open-source, unofficial toolkit for cauciones (overnight repo lending) on InvertirOnline, Argentina's largest retail broker: a Spanish CLI, a caución bot that runs in dry-run by default and needs a double opt-in to trade, a local MCP server so any coding agent (Claude Code, Codex, Cursor, Gemini) can drive it with propose → apply orders, and a macOS menu bar with a SwiftBar plugin and a native Swift app.

  • Native macOS menu bar app in SwiftUI (MenuBarExtra panel, light and dark, refreshing every 15 s) plus a SwiftBar plugin, both reading a status file so opening them never calls the broker API.
  • Local MCP server (stdlib JSON-RPC over stdio) with two-step writes: every order is proposed with a preview and a single-use id that expires in 120 s, and only applies when live mode is enabled.
  • Safety by default: dry-run everywhere, double opt-in for real orders, closed list of API endpoints, circuit breaker, per-order and daily limits; 73 tests and CI on macOS and Ubuntu (Python 3.11–3.13, Swift build).

langgraph-langserve-starter

2026

LangGraph · LangServe · FastAPI · Python

Production-grade REST API template exposing a LangGraph agent via LangServe. Multi-node research agent with SqliteSaver thread-resume, streaming, Langfuse observability, zero-key fake-LLM mode.

local-rag-chatbot

2026

LangGraph · RAG · Chroma · Ollama

Fully-local RAG chatbot built with LangGraph + Chroma + Ollama. Multi-Query Fusion and RRF retrieval, SqliteSaver chat history, pluggable LLM provider. Zero API keys required.

langgraph-ts-agents

2026

LangGraph · TypeScript · Multi-Agent · Zod

TypeScript port of LangGraph supervisor/worker patterns - streaming-first, fully typed with Annotation.Root + Zod, HITL approval, side-by-side Python references.

Somnio AI Tools

2026 - Present

CLI · AI Workflows · Python · Dart · Shell

Open-source CLI that installs and runs AI agent skills, health checks, and security audits for Flutter and NestJS projects across Claude Code, Cursor, Gemini, and other AI coding CLIs.

  • Added the `somnio run` command with interactive model selection, pre-flight validation, and per-step token usage tracking.
  • Refactored the CLI to a data-driven multi-agent registry (17 agents) and to an Agent Skills plugin architecture with Claude Code plugin support.
  • Built a custom workflow system for multi-step, multi-model AI task orchestration, plus Cursor CLI support and a setup command for AI CLI installation.

NovaMX CRM

2025

Next.js · Supabase · Edge Functions · OpenAI · LLMs

Real estate CRM with an LLM-powered lead/property classification engine. Public listings + internal lead tracker; agentic scoring runs on Edge Functions on top of Supabase.

Delivery Risk Maps

2026

MapLibre GL · OpenStreetMap · Node.js · Vercel Functions · SEO

Live risk maps for delivery riders in Córdoba and Buenos Aires, built on rider-reported zones and incidents from collaborative Google My Maps. Navigation-style MapLibre map with a 1–5 risk scale, report density, "how risky is this spot?" from your GPS, address search and night mode — plus server-rendered city and per-neighbourhood pages, sitemaps, JSON-LD and llms.txt so people searching for dangerous areas can find it.

Rent Calculator

2026

Next.js · React · TypeScript · Vitest · BCRA API · INDEC · SEO

Free rent tool for Argentina: updates a lease by ICL, IPC, Casa Propia, CAC, CER, UVA, IS or IPIM with official data from the BCRA, INDEC and CAMARCO (daily cache plus a committed fallback), showing every adjustment period, the next update and which values are still estimates. A move-or-stay comparison projects the rent with any index against the cost of moving. Formulas are pure functions covered by tests checked against the BCRA's own figures.

Stores Helper

2026

CLI · MCP · AI Agents · Play Store · App Store

MCP-style CLI exposing Play Store & App Store management (products, forms, subscriptions) as tools for AI agents. Lets coding agents read/write store data from the repo.

Spec Dev Ralph

2026

Spec Driven Dev

Starter template and guide for Spec Driven Development. Quick entry point for learning and applying the methodology.

Professional Experience

Independent AI & Full-Stack Engineer — Freelance

06/2025 – Present

Remote

  • Multi-tenant conversational AI agents in a TypeScript monorepo: Hono gateway, LangGraph.js with Postgres checkpointing, an MCP server, Langfuse/OpenTelemetry tracing, and automated evals.
  • Production AI agent products on Google Vertex AI (Gemini) with LangGraph.js and generative UI, exposed through a remote MCP server with its own OAuth 2.1 authorization server (Dynamic Client Registration, PKCE, token rotation), an evals harness with simulated users, Langfuse/OpenTelemetry tracing, a NestJS backend on Firebase Cloud Functions with serverless PDF rendering, and Flutter web apps.
  • Real-time, event-driven data systems in Python (asyncio, FastAPI, WebSockets) with Rust low-latency components, ingestion and replay pipelines (DuckDB, Parquet), and services operated on Linux under systemd.
  • Applied ML research on that data: fine-tuned a time-series foundation model (IBM Granite TSPulse) on GPU with PyTorch, and trained decision policies with PPO and evolution strategies in custom simulation environments, run as self-terminating AWS Spot jobs with datasets on Hugging Face.
  • 6+ white-label Flutter apps shipped from shared codebases: multi-flavor builds, Bloc/Riverpod, RevenueCat subscriptions, Firebase and Supabase, CI/CD with Codemagic and fastlane.
  • Next.js + Supabase products for small businesses: row-level security, Edge Functions, MercadoPago payments with webhooks, PWAs, PDF/Excel reporting, and ETL jobs.

FlutterFlow Developer / AI Native Engineer — Somnio Software

10/2024 – Present

Remote

  • Software engineering on internal tooling and platform work, using Python, TypeScript and Dart.
  • AI-assisted, spec-driven development practices within a product engineering team.
  • Contributor to the open-source Somnio AI Tools CLI.
  • Core contributor to FlutterFlow’s public issue triage (600+ issues).

Software Engineer — Detecta

07/2023 – 10/2024

Mexico (Remote)

  • Built and shipped the Detecta Security Platform, a triple-platform product serving internal operations, customers, and field drivers.
  • Integrated real-time GPS fleet tracking via Wialon, backed by Firebase and AWS microservices built with NestJS.

LowCode Specialist — Detecta

08/2022 – 07/2023

Mexico (Remote)

  • Delivered production mobile applications with FlutterFlow and Flutter on a Firebase backend.

IT Manager — California Fitness

10/2017 – 11/2021

Cancún, Mexico

  • Built and maintained a fully custom web CRM in PHP/MySQL, shipping weekly improvements across all branches.
  • Introduced QA automation with Selenium and Cypress covering the CRM and mobile applications.
  • Replaced manual business processes with automated workflows; maintained and optimized the production database.
  • Managed the external ERP integration project with the vendor, and ran eCommerce and online client support.

Branch Office Manager — California Fitness

01/2016 – 10/2017

Cancún, Mexico

  • Managed branch operations, staff training, inventory, and provider relationships before moving into the IT organization.

Certifications

  • IBM Generative AI Engineering — Professional Certificate, IBM via Coursera, 10/2026 (16 courses)
    • Courses: AI Agents Using RAG and LangChain; Generative AI Applications Specialist; Generative AI Advanced Fine-Tuning for LLMs; Generative AI Engineering with Transformers & LLMs; Deep Learning Essentials with Keras; Generative AI and LLMs: Architecture and Data Preparation; Generative AI Foundational Models for NLP & Language Understanding; Generative AI Language Modeling with Transformers; Data Analysis with Python; Machine Learning with Python (V2); Python Project for AI and Application Development; Building Generative AI-Powered Applications with Python; Artificial Intelligence Essentials V2; Generative AI Essentials; Generative AI: Prompt Engineering
  • Agentic AI Engineering — Specialization, Edureka via Coursera, 05/2026 (4 courses)
    • Courses: Developing MCP-Powered Agentic AI Systems; Multi-Agent Systems with LangGraph; Applied Agentic AI Pipelines with LangChain; Building Your First AI Agent with LangChain
  • AWS Certified Cloud Practitioner — Certification, Amazon Web Services, 05/2024

Education

Engineering, Systems Engineering — Universidad Interamericana para el Desarrollo, Mexico

2014 – 2017

Diploma, Mobile Application Development — Deusto Formación, Mexico

2015

Bachelor's, International Business — Universidad Tecmilenio, Mexico

2011 – 2013