# Matias Lapolla

> AI Engineer · Agent platforms, evals, MCP · TypeScript, Python. Tandil, Argentina (GMT-3) · Remote.

Hi, I’m Matias Lapolla. I build AI-native products at Persiscal, mostly agent platforms. On the side I run [Senda](https://trysenda.xyz/), where I ship software experiments with AI agents. Each one starts by defining its acceptance criteria, because I believe [an agent is only as good as the evals you run on it](https://matiaslapolla.com/blog/build-ai-agents-start-with-evals). I write about building with agents: [why the tool description matters more than the code](https://matiaslapolla.com/blog/mcp-the-tool-description-is-the-interface) and [what memory is for](https://matiaslapolla.com/blog/mastra-memory). My [CV](https://matiaslapolla.com/cv) has the rest, the [blog](https://matiaslapolla.com/blog) has every article, and you can reach me on [X](https://x.com/matiasdev_).

The CV: [HTML](https://matiaslapolla.com/cv) · [Markdown](https://matiaslapolla.com/cv.md).

## Work

- **Tuxdi** · UI/UX & Graphic Design · 2020–2021
- **GrupoAgni** · Full Stack · 2021–2024
  - Government Tech
  - PYME Software
- **Persiscal** · Full Stack, AI Engineer · 2024– (current)
  - AI Agents Platform · 2026–
  - GeoScan · 2026–
  - Starlings Travel · 2024–2026

## On the side

- **Youpy** · CTO · 2026–: Proptech startup: CRM, property search and a WhatsApp AI agent.
- **[Senda](https://trysenda.xyz/)** · Software experiments: Production agents you own.

## Open source

All repos: https://github.com/matiaslapolla

- [agentsi](https://github.com/matiaslapolla/agentsi) (Rust · TS): A desktop app and Claude Code plugin where role agents (PM, dev, QA, scout, social) rank the work and the owner approves each change.
- [langgraph](https://github.com/matiaslapolla/langgraph) (Python): A LangGraph lab: human-in-the-loop approval, memory, local llama.cpp models and evals with graders, a baseline and Langfuse.
- [mini-ai-sandbox](https://github.com/matiaslapolla/mini-ai-sandbox) (TypeScript): A Docker execution workspace for Mastra agents over HTTP, SSE and MCP.
- [portwatch](https://github.com/matiaslapolla/portwatch) (TypeScript): A TUI, CLI and MCP server that lets people and coding agents find and free local ports, aware of git worktrees.
- [mdnm](https://github.com/matiaslapolla/mdnm) (Rust · Tauri): A desktop Markdown viewer with first-class Mermaid diagrams.
- [readaloud](https://github.com/matiaslapolla/readaloud) (Python): Terminal TTS that reads the last Claude Code reply aloud with a local Kokoro voice.

## Writing

- 2026-10-20 · [SOLID for AI Agents: The Charter Is the Interface, and KISS Decides When to Split](https://matiaslapolla.com/blog/solid-and-kiss-for-ai-agents): Single responsibility, interface segregation, open/closed, Liskov and dependency inversion applied to a LangGraph agent split into four specialists: what each principle changed in the design, which contracts broke, and when KISS says one agent is enough.
- 2026-09-02 · [So You Want to Build AI Agents That Do the Work? Start with Evals](https://matiaslapolla.com/blog/build-ai-agents-start-with-evals): Building an agent in LangGraph takes an afternoon. Knowing how often it does the work is the job. The graph, the eval suite that measured it, a seven-search spiral the numbers exposed, a bake-off across seven local models, and what the one-line fix quietly cost.
- 2026-07-03 · [Mastra Memory: The Context Layer That Survives Long Conversations](https://matiaslapolla.com/blog/mastra-memory): Working memory, semantic recall, and Observational Memory in Mastra — why the last-N-messages window leaks, how the three recall tiers stack, and the three-agent loop that keeps agents sharp past 200 turns.
- 2026-06-08 · [MCP: Why the Description String Is the Most Important String in Your Harness](https://matiaslapolla.com/blog/mcp-the-tool-description-is-the-interface): The model never reads your tool code — it reads the description string and programs against it. A field report on writing tool contracts for a Mastra support agent, exposing them over MCP, and why that one string is both your interface and your attack surface.
- 2026-06-01 · [Mastra Processors: la capa de guardrails que se ejecuta en cada paso](https://matiaslapolla.com/blog/mastra-processors-and-guardrails-es): Un recorrido por los processors de Mastra — cómo funcionan los pipelines de entrada y salida, cuándo abortar y cuándo reintentar, y cómo construir guardrails que aguanten en producción.
- 2026-06-01 · [Mastra Processors: The Guardrail Layer That Runs on Every Step](https://matiaslapolla.com/blog/mastra-processors-and-guardrails): A deep dive into Mastra processors — how the input/output pipelines work, when to abort vs. retry, and how to build guardrails that actually hold in production.
- 2026-05-12 · [Building a Fully Functional Agent Harness with Mastra](https://matiaslapolla.com/blog/building-a-fully-functional-harness-with-mastra): A practical walkthrough of harness engineering in 2026: mapping the agent loop, tool interface, permission system, processors, memory, and evals to concrete Mastra + Next.js primitives.
- 2026-01-24 · [Foundations of AI: From Vectors to Neural Networks](https://matiaslapolla.com/blog/from-vectors-to-networks): A comprehensive guide to the mathematical foundations of modern artificial intelligence, connecting linear algebra, probability, calculus, and neural networks to attention mechanisms and transformers
- 2025-12-27 · [Vibe Engineering: A Guide to AI-Assisted Coding with Agents](https://matiaslapolla.com/blog/vibe-engineering-coding-in-the-era-of-agents): Learn proven techniques for AI-assisted coding with Claude Code, Cursor, and Codex. Includes context engineering strategies, task atomization patterns, and practical tips for maximizing AI agent effectiveness.
- 2025-07-26 · [How I Use AI and LLMs for Coding - A Comprehensive and Opinionated Guide](https://matiaslapolla.com/blog/how-i-use-ai-for-coding): An overview of how I use AI and LLMs for my coding tasks. This is not a mandatory guide to follow but this article can help you get started with AI in your coding workflow.
- 2025-04-19 · [Model Context Protocol (MCP) - The USB-C for AI](https://matiaslapolla.com/blog/mcp-the-usb-c-for-ai): An overview of the Model Context Protocol (MCP), a new standard for AI model communication, the Missing Abstraction Layer between LLMs and the Real World

## Contact

- Email: hi@matiaslapolla.com
- X: https://x.com/matiasdev_
- GitHub: https://github.com/matiaslapolla
- CV: https://matiaslapolla.com/cv
- Blog: https://matiaslapolla.com/blog
- LinkedIn: https://www.linkedin.com/in/matiaslapolla
